AI Coding Plan

Coding Plan

Kimi

Kimi Membership | Kimi Code

• Tiers (RMB/month): Andante 49, Moderato 99, Allegretto 199, Allegro 699
• Subscription renewal: due to tight compute resources, priority is given to currently subscribed users; renewal was available until 2026-08-20 00:00, after which direct purchase is no longer available
• Kimi Code is powered by the flagship K3 model (~2.8 trillion parameters) and offers the K2.8 Preview standard mode and the K2.7 Code HighSpeed mode, with a maximum inference speed of 260 Tokens/s and a 1M Token context window; it is compatible with mainstream agent tools including Kimi Code CLI, Claude Code and VS Code
• Model IDs and membership tier requirements: k3 and k3-256k require Moderato / Plus or above, with 1M context requiring Allegretto / Pro or above; kimi-for-coding (K2.8 Preview) requires Andante / Plus or above; kimi-for-coding-highspeed requires Allegretto / Pro or above
• API endpoints are split into domestic and overseas: OpenAI-compatible (both the Chat Completions and the Responses protocols; the documented Responses endpoint is the domestic https://api.kimi.com/coding/v1/responses) — domestic https://api.kimi.com/coding/v1, overseas https://api.kimi.ai/coding/v1; Anthropic-compatible — domestic https://api.kimi.com/coding/, overseas https://api.kimi.ai/coding/; Kimi Open Platform — domestic https://api.moonshot.cn/v1, overseas https://api.moonshot.ai/v1
• API pricing (RMB per 1M tokens):
  • K3: cached input 2, uncached input 20, output 100, 1M context
  • K2.7-Code: cached input 1.3, uncached input 6.5, output 27, 256K context
  • K2.7-Code-HighSpeed: cached input 2.6, uncached input 13.0, output 54, 256K context

MiniMax

MiniMax M Plan | FAQ | International

• M Plan is MiniMax's individual AI subscription and replaces the former Token Plan (Token Plan is no longer sold). Monthly list price: Go ¥49 / $22, Explore ¥119 / $55, Build ¥469 / $132; annual billing costs 10 months (2 months free); the first month is 50% off during the launch (through 2026-10-14)
• Included models: all tiers include the text model MiniMax-M3.1-Flash-Preview (natively multimodal, 1M context, tunable thinking depth), image and audio generation, web search and multimodal understanding; Explore and Build additionally include the H3 video model (Go excludes video); music and similar resources are not included
• Usage (relative to Go): Explore is 3x; Build is 12x in the Chinese-site docs but 7.5x in the international docs (the two sites disagree); text/image/audio and other non-video models are subject to 5-hour and weekly windows, while video models are subject only to the weekly window
• Credit packs: 1,000 credits = ¥7 (international $1), valid for 1 year from purchase; plan quota is used before credit packs
• The official docs state MiniMax-M3.1-Flash-Preview is available only through M Plan and MiniMax Code for now
• Model retirement notice: the MiniMax M2.5 and M2.7 families are being phased out — from 2026-10-19 they are no longer available to new users, and on 2026-11-18 they are retired; existing pay-as-you-go API users can keep calling them until the retirement date; for M Plan and Token Plan users, requests to M2.5 and M2.7 are routed to M3.1 Flash Preview from 2026-10-19 (the original model names are kept as compatible entry points, so no model names need changing, and plan usage entitlements are unchanged)
• The Subscription Key works with any tool that speaks the OpenAI- or Anthropic-compatible protocol (the docs list Claude Code, Codex, Cursor, OpenCode, Pi, Hermes Agent and OpenClaw); using MiniMax Code requires no separate subscription or Subscription Key; tiers cannot be downgraded — monthly plans can be upgraded to a higher monthly or annual tier, and annual plans only to a higher annual tier; once auto-renewal is turned off it cannot be restored (buy again; value remaining before expiry can be credited toward an upgrade)
• Existing Token Plan subscribers: with auto-renewal on, the current plan and benefits stay unchanged (retaining historical perks such as unlimited quotas and the 150% weekly quota), and you can upgrade to M Plan at any time (the unused portion is prorated and credited); upgrade mapping: Plus → Go/Explore/Build, Max → Explore/Build, Ultra → Build; after upgrading you cannot return to Token Plan
• Legacy Token Plan migration: in China, Max-hs ¥199 → Max ¥119 (fee down ¥80 plus ~¥160 of credits granted monthly) and Ultra-hs ¥899 → Ultra ¥469 (down ¥430 plus ~¥860 of credits granted monthly), with Starter ¥29 and Plus-hs ¥98 retained for existing subscribers only; internationally, Max-hs $80 → Max and Ultra-hs $150 → Ultra (each down \(30 plus ~\)60 of credits granted monthly), with Starter $10 and Plus-hs $40 retained; annual plans of retired tiers receive equivalent credits monthly (each grant valid for 1 year, non-rolling)
• Legacy benefits: benefits for users who subscribed before 2026-06-05 are retained; compensation credits are valid for 1 year; entitlement bonuses remain valid only during continuous subscription periods and are waived when changing tiers or cancelling
• MiniMax M3 API pricing (RMB per 1M tokens):
  • <= 512K input tokens: cached input 0.42, uncached input 2.10, output 8.40
  • > 512K input tokens: cached input 0.84, uncached input 4.20, output 16.80
  • 1M context

Zhipu (GLM, Z.ai)

Zhipu GLM Coding Plan

• Lite Plan (118 RMB/month): 2,000 points per 5 hours, 10,000 points per week
• Pro Plan (538 RMB/month): 12,000 points per 5 hours, 60,000 points per week
• Max Plan (1078 RMB/month): 28,000 points per 5 hours, 140,000 points per week
• Point refresh rules: the 5-hour points refresh dynamically (the allowance resets 5 hours after consumption) and the weekly points refresh on a 7-day cycle starting from the time the plan was purchased; after a subscriber upgrades, the points allowance refreshes immediately to the new tier, whereas renewal does not trigger a points reset and the existing refresh cycle is unchanged
• Model points consumed = (input tokens × Input coefficient + cached input tokens × Cached Input coefficient + output tokens × Output coefficient) / 10,000
• MCP points consumed = number of calls × Output coefficient
• All plans support GLM-5.3 and GLM-5.3-Flash; requests for the previous models GLM-5.2/GLM-5.1 are automatically routed to GLM-5.3, and requests for GLM-5-Turbo/GLM-4.7 are automatically routed to GLM-5.3-Flash
• Off-peak hours consume points at 50% of the base rate; peak hours are weekdays 14:00–18:00 (UTC+8)
• Limited-time campaigns (both ended): the "GLM-5.3-Flash Usage Campaign" (2026-09-03 to 2026-10-07, daily 23:00–09:00, unlimited GLM-5.3-Flash usage via ZCode and AutoClaw plus doubled quota on other agents) and the "Double Festival" campaign (2026-09-25 to 2026-10-07, all-day usage charged at the off-peak rate) expired on 2026-10-07; as of 2026-10-09 no limited-time campaign is running on either site
• Coefficients: GLM-5.3 — Input 6.9 / Cached Input 1.7 / Output 24; GLM-5.3-Flash (including the visual understanding MCP) — Input 2.3 / Cached Input 0.56 / Output 8
• Token usage varies with the cache hit rate:

Cache Hit Rate	Model	Lite (M Tokens/week)	Pro (M Tokens/week)	Max (M Tokens/week)
95%	GLM-5.3	0.48～0.97	2.90～5.80	6.76～13.52
95%	GLM-5.3-Flash	1.46～2.92	8.77～17.55	20.47～40.95
96%	GLM-5.3	0.50～0.99	2.97～5.95	6.94～13.87
96%	GLM-5.3-Flash	1.50～3.00	9.00～18.01	21.01～42.02
98%	GLM-5.3	0.52～1.04	3.13～6.27	7.31～14.63
98%	GLM-5.3-Flash	1.58～3.17	9.50～19.00	22.17～44.33

Range explanation: maximum tokens assumes all usage in off-peak hours at 0.5× points; minimum tokens assumes all usage in peak hours at 1×. By fully utilizing the off-peak discount, you can save up to 92% compared with pay-as-you-go calls to the GLM-5.3 standard API - API pricing (RMB per 1M tokens): - GLM-5.3: cached input 2, uncached input 8, output 28, 1M context - GLM-5.3-Flash: cached input 0.23, uncached input 0.8, output 2.8, 1M context - GLM-5.2: cached input 2, uncached input 8, output 28, 1M context - GLM-5.1: cached input 1.3/2, uncached input 6/8, output 24/28, 200K context - GLM-5-Turbo: cached input 1.2/1.8, uncached input 5/7, output 22/26, 200K context - GLM-4.7: cached input 0.4/0.6/0.8, uncached input 2/3/4, output 8/14/16, 200K context

Zhipu GLM Coding Plan Team Edition

• Team Standard (598 RMB/month): maximum 60M tokens per seat per 5 hours, maximum 300M tokens per seat per week
• Team Advanced (1198 RMB/month): maximum 160M tokens per seat per 5 hours, maximum 800M tokens per seat per week
• "Maximum" refers to the total tokens that can actually be consumed at a 1× consumption coefficient; current per-model consumption rules:
  • GLM-4.7, GLM-4.5-Air: consume at 1× all day
  • GLM-5.2, GLM-5-Turbo: 3× during peak hours and 2× during off-peak hours; as a limited-time benefit, off-peak hours consume only 1×
  • Peak hours are 14:00–18:00 (UTC+8) daily

Zhipu International GLM Coding Plan: all plans support GLM-5.3 and GLM-5.3-Flash; requests for GLM-5.2/GLM-5.1 auto-route to GLM-5.3, and GLM-4.7 auto-routes to GLM-5.3-Flash

Cloud Providers

• Volcano Engine Coding Plan (Personal Edition) | Documentation
  • Lite Plan (40 RMB/month): maximum ~1,200 requests per 5 hours, ~9,000 per week, ~18,000 per subscription month
  • Pro Plan (200 RMB/month): 5x the Lite Plan quota
  • Supported models: Doubao-Seed-2.1-pro, Doubao-Seed-2.1-lite, Doubao-Seed-2.0-mini, Doubao-Seed-Evolving, MiniMax-M3, Kimi-K2.7-Code (being retired), Kimi-K2.8-Preview, Kimi-K3 (high deduction coefficient, recommended only for Pro plan users), GLM-5.3, GLM-5.3-Flash, DeepSeek-V4-Flash, DeepSeek-V4-Pro, DeepSeek-V4.1-Flash
  • 1M-context support: doubao-seed-evolving, glm-5.3, glm-5.3-flash, kimi-k3, kimi-k2.8-preview, deepseek-v4.1-flash, deepseek-v4-flash, deepseek-v4-pro
  • Limited-time promotions: deepseek-v4.1-flash deduction coefficient at 50% off (2026-09-23 00:00 to 2026-12-30 23:59:59); Kimi-K2.8-Preview quota campaign (2026-09-18 00:00 to 2026-12-14 23:59:59, comparable to the Agent Plan 40%-off promotion period)
  • Key model specs: Doubao-Seed-2.1-pro (new-generation flagship, 1024k context / 256k output), Doubao-Seed-2.1-lite (lightweight and efficient, 1024k/256k), Doubao-Seed-2.0-mini (fast responses, 256k/128k), DeepSeek-V4.1-Flash (552B total-parameter MoE, native multimodal vision understanding, 1024k/384k), Kimi-K2.8-Preview (close to K3 with higher thinking efficiency, 1M/1M, text and image input), GLM-5.3-Flash (Zhipu's first natively multimodal model, 320B total params / 18B active, 1M/128K)
• Volcano Engine Agent Plan (Personal Edition)
  • Agent Fuel Points (AFP) are the unified billing unit: text/embedding models = (input token × input coefficient + output token × output coefficient) / 10,000; video models = tokens / 10,000 × coefficient; image models = number of successfully generated images × coefficient. Coefficients are determined solely by the model and do not vary with input length
  • Deduction coefficients per model (input/output identical): doubao-seed-2.0-mini 0.25 (without audio input) / 2.5 (with audio input); deepseek-v4-flash 0.5; doubao-seed-2.1-lite 0.5 (4.5 with audio input); glm-5.3-flash 0.5; doubao-seed-evolving, minimax-m3, doubao-seed-2.1-pro 2.5; deepseek-v4.1-flash 2.5 (1.25 at a limited-time 50% off, 2026-09-15 00:00 to 2026-12-30 23:59:59); kimi-k2.7-code (being retired) 4.5; kimi-k2.8-preview 8 (4.8 at a limited-time 40% off, 2026-09-17 00:00 to 2026-12-14 23:59:59); glm-5.3 (glm-latest) 4.5; deepseek-v4-pro 5.5; kimi-k3 10 (5 at a limited-time 50% off, 2026-10-12 00:00 to 2026-10-25 23:59); doubao-embedding-vision 0.5
  • Tiers (AFP, per 5 hours / week / month / day): Small 40 RMB/month (2000/7000/20000/10000), Medium 200 (10000/35000/100000/50000), Large 500 (25000/87500/250000/125000), Max 1000 (50000/175000/500000/250000)
  • Image, video and speech models and Harness have no 5-hour or weekly limits; they are subject only to the daily and monthly limits, with the daily limit equal to half the monthly limit
  • All plans support: doubao-seed-2.1-pro, doubao-seed-2.1-lite, doubao-seed-2.0-mini, deepseek-v4-flash, deepseek-v3.2, minimax-m3, glm-5.3, glm-5.3-flash, kimi-k2.7-code (being retired), kimi-k2.8-preview, deepseek-v4-pro, deepseek-v4.1-flash, doubao-embedding-vision, doubao-seedream-5-0-pro, doubao-seed-tts-2.0, doubao-seed-asr-2.0
  • Medium and above additionally support: doubao-seedance-2.0, doubao-seedance-2.0-fast, doubao-seedance-2.0-mini, doubao-seedance-2.5
  • Agent Evolution: first 50 files free
• Alibaba Cloud Bailian Token Plan (Personal Edition)
  • Lite (60, 39 for a limited time), Essential (120, 79), Standard (180, 139), Pro (600, 499) RMB/month, corresponding to 11,500 / 25,500 / 45,000 / 180,000 Credits per month; Usage Pack (100 RMB/month) = 20,000 Credits
  • Since 2026-09-22 the fixed 7-day window quota has been abolished in favour of a subscription-month quota (from 2026-10-14 the official wording is "the quota is counted by subscription month", resetting based on the subscription date rather than a fixed calendar date); once cumulative consumption in the subscription month reaches the plan quota, service is paused, and unused quota does not roll over
  • Deduction order: each request is deducted first from the current subscription month's plan quota, and once exhausted from the Usage Pack quota
  • Upgrades: the current subscription cycle is not reset; the price difference is paid for the remaining days and the new quota is granted for the current cycle, with the new tier's monthly quota applying from the next subscription month
  • Essential supports 2–3 concurrent agents and carries all Lite benefits
  • For the subscriber's own use on a single device only
  • Supported models: auto, qwen3.8-max, qwen3.8-flash, qwen3.7-max, qwen3.7-plus, qwen3.6-flash, qwen-image-3.0-pro, qwen-audio-3.0-tts-plus, qwen-audio-3.0-realtime-plus, qwen-audio-3.0-asr-flash, wan2.7-image, wan2.7-image-pro, wan3.0-video, deepseek-v4.1-flash, deepseek-v4-pro, deepseek-v4-pro-0813, deepseek-v4-flash-0731, glm-5.3, glm-5.2, happyhorse-1.1-i2v, happyhorse-1.1-t2v, happyhorse-1.1-r2v, decision-model-preview
  • Limited-time night discount (22:00–08:00 daily): 40% off Credits for qwen3.8-max and qwen3.8-flash; 50% off for deepseek-v4-pro-0813, deepseek-v4-flash-0731 and deepseek-v4.1-flash
  • decision-model-preview (domain model – decision model) is free for a limited time and consumes no Credits
• Alibaba Cloud Bailian Token Plan (Team Edition)
  • Standard Seat ¥198/seat/month (25,000 Credits), Advanced Seat ¥698 (100,000), Premium Seat ¥1,398 (250,000); Shared Usage Pack ¥5,000/pack (625,000 Credits)
  • Credits consumed per request are dynamically determined by model type, token usage, reasoning mode and tool calls; for example, with Qwen3.6-plus every 5,000 uncached input tokens, every 50,000 cached input tokens or every 5,000/6 output tokens equals one Credit
  • API price per Credit (implicit caching): 0.01–0.02 RMB within 256K context, 0.04–0.08 RMB for 256K–1M
  • Supported models: auto, qwen3.8-max, qwen3.8-flash, qwen3.7-max, qwen3.7-plus, qwen3.6-plus, qwen3.6-flash, qwen-image-2.0, qwen-image-2.0-pro, qwen-image-3.0-pro, qwen-audio-3.0-tts-plus, qwen-audio-3.0-realtime-plus, qwen-audio-3.0-asr-flash, wan2.7-image, wan2.7-image-pro, deepseek-v4.1-flash, deepseek-v4-pro, deepseek-v4-pro-0813, deepseek-v4-flash-0731, kimi-k2.7-code, kimi-k2.6, kimi-k2.5, glm-5.3, glm-5.2, glm-5.1, glm-5, MiniMax-M2.5, happyhorse-1.1-i2v, happyhorse-1.1-t2v, happyhorse-1.1-r2v
• Tencent Cloud LLM Token Plan
  • Enterprise Professional Plan: 1 RMB/100 Credits, minimum purchase 50K Credits (500 RMB/month); Light Enjoyment Plan: 2 RMB per million tokens
  • Enterprise supported models (varies slightly by region): Auto, Hy4 preview, GLM-5.3, GLM-5.3-Flash, GLM-5.2, GLM-5 (retiring 2026-10-09), GLM-5.1 (retiring 2026-10-09), GLM-5-Turbo (retiring 2026-10-09), Kimi K2.7 Code, Kimi K2.7 Code HighSpeed, Kimi K3, Kimi-K2.6, MiniMax-M2.7, MiniMax-M3, DeepSeek-V4-Pro, DeepSeek-V4-Flash 0731 Official, DeepSeek-V4-Pro 0813 Official, DeepSeek-V4.1-Flash (not Official direct-supply), DeepSeek-V4.1-Flash Official direct-supply, DeepSeek-V4-Flash 0731 Official direct-supply, DeepSeek-V4-Pro 0813 Official direct-supply, DeepSeek-V4-Flash-Vision-Exp Official direct-supply, MiMo-V2.6-Pro, MiMo-V2.6-Flash
  • Peak-valley billing: for DeepSeek models, peak hours are weekdays (Mon–Fri) 9:00–12:00 and 14:00–18:00 with all other times off-peak; weekends are billed at off-peak rates all day
  • Credit prices (credits per 1M tokens): DeepSeek-V4.1-Flash (including Official direct-supply) cached 2 / uncached 100 / output 400 (off-peak), 4 / 200 / 800 (peak); Hy4 preview 30 / 600 / 1800; MiMo-V2.6-Pro 2.5/300/600 in Guangzhou and 2.59/313.05/626.1 in Singapore; MiMo-V2.6-Flash 2/100/200 in Guangzhou and 2.02/100.75/201.5 in Singapore
  • Personal Hy Token Plan: Lite 28 RMB/month (560 credits), Standard 78 (1,560), Pro 238 (4,760), Max 468 (9,360); supports Hy3 and Hy4 preview
  • Personal Universal Token Plan: Lite 39 (780 credits), Standard 99 (1,980), Pro 299 (5,980), Max 599 (11,980); supports Auto, DeepSeek-V4.1-Flash Official direct-supply, DeepSeek-V4-Flash Official direct-supply, DeepSeek-V4-Pro Official direct-supply, MiniMax-M2.7, MiniMax-M3, GLM-5, GLM-5.1, GLM-5.2, GLM-5.3, GLM-5.3-Flash, Kimi K2.7 Code, Kimi K3, Hy4 preview, MiMo-V2.6-Flash
• Baidu Qianfan Token Plan (Personal Edition) | Personal Edition Documentation | Enterprise Edition Documentation
  • Mini 9.9 RMB/month (10M tokens), Lite 40 (42M), Pro 200 (230M), Max 600 (700M)
  • Supported models: DeepSeek-V4-Pro, DeepSeek-V4-Flash, GLM-5.2, GLM-5.1, Kimi-K2.6, ERNIE 5.1
• Baidu Qianfan Coding Plan
  • Coding Plan Lite at 40 RMB/month: up to ~1,200 requests per 5 hours, ~9,000 per week, ~18,000 per subscription month; Coding Plan Pro at 200 RMB/month: 6,000 / 45,000 / 90,000
  • Supported models and deduction coefficients: DeepSeek-V4-Pro (peak 5 / off-peak 4), GLM-5.1 (peak 4 / off-peak 3), Baidu Search (10); peak hours are generally 10:30-12:00 and 14:00-18:00 daily (varies with traffic)
  • Official notice: "the supported model range will be adjusted dynamically with model iterations. Coding Plan has stopped model updates; users who need newer models are advised to migrate to Token Plan (Personal Edition)". This update reduced the model library to the two language models above (the Model Names that can be set in the config file likewise dropped to deepseek-v4-pro and glm-5.1), removing Kimi-K2.5, DeepSeek-V3.2, DeepSeek-V4-Flash, GLM-5, MiniMax-M2.5 and ERNIE-4.5-Turbo-20260402
• JD Cloud Coding Plan
  • Lite Plan (19.9 first purchase, 40 renewal, RMB/month): maximum 1,200 requests per 5 hours, 9,000 per week, 18,000 per subscription month
  • Pro Plan (99.9 first purchase, 200 renewal, RMB/month): maximum 6,000 requests per 5 hours, 45,000 per week, 90,000 per subscription month
  • Supported models: DeepSeek-V3.2, GLM-5, GLM-4.7, MiniMax-M2.5, Kimi-K2.5, Kimi-K2-Turbo, Qwen3-Coder
• iFlytek Astron Token Plan Team Edition | Subscription
  • Standard Member 200 RMB/seat/month (20,000 Credits, 2M TPM), Advanced Member 600 (60,000 Credits, 3M TPM), Premium Member 1200 (200,000 Credits, 5M TPM)
  • Supported models: Spark-X2.5, Spark-X2-Flash, GLM-5.2, GLM-5.1, GLM-5, DeepSeek-V4-Pro, DeepSeek-V4-Flash, DeepSeek-V3.2, Kimi-K2.6, Kimi-K2.5, MiniMax-M2.5, Qwen3.5-397B-A17B, Qwen3.6-35B-A3B, Qwen3.5-35B-A3B, Qwen3-Coder-Next-FP8, GLM-4.7-Flash
• iFlytek Astron Coding Plan | Subscription
  • Professional (39 RMB/month): ~1,200 requests per 5 hours, ~9,000 per week, ~18,000 per subscription month; supports Spark-X2-Agent, Spark-X2, Auto, GLM-5.1, GLM-5, MiniMax-M2.5, Kimi-K2.6, Kimi-K2.5, DeepSeek-V3.2, Spark-X2-Flash, Qwen3.6-35B-A3B, GLM-4.7-Flash, Qwen3.5-35B-A3B, Qwen3-Coder-Next-FP8, Qwen3.5-397B-A17B
  • Efficient (199 RMB/month): ~6,000 requests per 5 hours, ~45,000 per week, ~90,000 per subscription month; supports Spark-X2-Agent, Spark-X2, Auto, GLM-5, GLM-5.2, DeepSeek-V4-Pro, DeepSeek-V4-Flash, MiniMax-M2.5, Kimi-K2.6, Kimi-K2.5, DeepSeek-V3.2, Spark-X2-Flash, Qwen3.6-35B-A3B, GLM-4.7-Flash, Qwen3.5-35B-A3B, Qwen3-Coder-Next-FP8, Qwen3.5-397B-A17B
• CTCloud Programming Token Plan (Credits Edition)
  • 29 (3,000 credits), 89 (10,000), 199 (25,000), 399 (50,000), 699 (100,000) RMB/month
  • Supported models: DeepSeek-V4-Pro, DeepSeek-V4-Flash-0731, GLM-5.2, GLM-5.1, Kimi-K2.6, MiniMax-M3
  • Credit-to-token conversion (tokens per credit, input/output): DeepSeek-V4-Pro 1,111/370; DeepSeek-V4-Flash 3,333/1,111; GLM-5.2 1,250/357; GLM-5.1 input [0,32k] 1,667 (output 417), (32k,200k] 1,250 (output 357); Kimi-K2.6 1,538/370; MiniMax-M3 input [0,512k] 4,762 (output 1,190), (512k,1M] 2,381 (output 595)
  • The previous token-metered plans allow renewals by existing subscribers only; since 2026-10-10 they support DeepSeek-V4-Flash-0731, GLM-5.1 and GLM-5.2 (GLM-5.0 and DeepSeek-V3.2 were removed as previously announced)
• Huawei Cloud MaaS Token Plan
  • Lite 59 RMB/month (50M tokens per subscription month), Standard 149 (130M), Pro 399 (380M), Max 799 (880M)
  • Supported models: GLM-5, GLM-5.1, Kimi-K2.6, DeepSeek-V3.2, DeepSeek-V4-Flash

Others

• StepFun Step Plan
  • Flash Mini 49 RMB/month (400M Credits), Flash Plus 99 (1600M), Flash Pro 199 (8000M), Flash Max 699 (40000M)
  • Supported models: step-5-preview (next-generation flagship base model), step-3.7-flash, step-3.5-flash-2603, step-3.5-flash, stepaudio-2.5-realtime, stepaudio-2.5-chat, stepaudio-2.5-tts, stepaudio-2.5-asr, step-router-v1 (intelligent routing between deepseek-v4-pro and step-3.5-flash)
• Xiaomi MiMo Token Plan (Personal Edition)
  • Lite (39 RMB or 6 USD/month): 4.1B Credits per month; Standard (99 RMB or 16 USD): 11B; Pro (329 RMB or 50 USD): 38B; Max (659 RMB or 100 USD): 82B
  • Supported models: MiMo-V2.6-Pro, MiMo-V2.6-Flash, MiMo-V2.5-Pro, MiMo-V2.5, MiMo-V2.5-ASR, MiMo-V2.5-TTS-VoiceClone, MiMo-V2.5-TTS-VoiceDesign, MiMo-V2.5-TTS
  • mimo-v2.5-pro and mimo-v2.5 will be retired at 10:00 Beijing time on 2026-10-21; switching to the new models as soon as possible is recommended
  • Discounts: 88% off first purchase, 88% off annual auto-renewal, 0.8× consumption at night (0:00–8:00)
• OpenCode Go
  • On 2026-10-13 the official docs page moved to v2: the old URL /docs/zh-cn/go (which returned 404 on 2026-10-12) now 302-redirects to https://opencode.ai/v2/docs/console/go/, so the Go content returns from the marketing page (https://opencode.ai/zh/go , which lists only 13 featured models) back to the docs site; the page now publishes token prices, monthly allowances, estimated requests (5 hours/week/month), per-request token estimates and data retention for all 32 models
  • Two plans: Go (\(10/month)** and **Go Plus (\)40/month); token prices are identical and Go Plus only raises the per-model usage limits (monthly allowances are 2–8x those of Go, mostly 4x)
  • Usage limits are defined per model as a monthly amount: 5 hours = 20% of the monthly limit, week = 50%, month = 100%
  • Endpoints: the docs now use https://opencode.ai/inference/go/v1/{chat/completions,messages,responses} (the former https://opencode.ai/zen/go/v1/… still works — the two domains are aliases and /v1/models returns 200 on both); reference models inside OpenCode as opencode-go/<model-id> (e.g. opencode-go/kimi-k3), and list them via https://opencode.ai/inference/go/v1/models
  • Monthly usage allowances (Go / Go Plus): $15 / $60 — GLM-5.3, Kimi K3, MiMo-V2.6-Pro, MiMo-V2.5-Pro, Qwen3.8 Max, DeepSeek V4 Pro, Grok 4.6, Grok 4.7, GPT 6 Luna, GPT 5.6 Luna, Claude Haiku 5.5; $30 / $90 — Qwen3.8 Flash; $30 / \(120 — Hy4 preview, Space Bunny; ~~\)30~~ \(60 / ~~\)120~~ $240 (2x · Limited time) — Gemini 4 Argon Contributor; $60 / $120 — MiMo-V2.6-Flash, MiMo-V2.5, Muse Spark 1.3 Contributor, Muse Spark 1.2 Contributor; $60 / $180 — GLM-5.3-Flash, GLM-5.2, Kimi K2.7 Code, MiniMax M3, Qwen3.7 Plus; $60 / $240 — Kimi K2.6, LongCat-2.0, MiniMax M2.7, Hy3; $80 / $160 — DeepSeek V4.1 Flash (raised on 2026-10-12; estimated 5-hour requests 34,670 / 69,330); unlimited for a limited time — LongCat 2.5 Preview Free, Step 5 Preview Free, Cat Jam Free
  • Token prices (identical on Go and Go Plus, per 1M tokens: input / output / cache read / cache write): GLM-5.3-Flash 0.15 / 0.50 / 0.03; GLM-5.3, GLM-5.2 1.40 / 4.40 / 0.26; Kimi K3 3.00 / 15.00 / 0.30; Kimi K2.7 Code 0.95 / 4.00 / 0.19; Kimi K2.6 0.95 / 4.00 / 0.16; LongCat-2.0 0.30 / 1.20 / 0.006; MiMo-V2.6-Flash, MiMo-V2.5 0.14 / 0.28 / 0.0028; MiMo-V2.6-Pro, MiMo-V2.5-Pro 0.435 / 0.87 / 0.003625; MiniMax M3 0.30 / 1.20 / 0.06; MiniMax M2.7 0.30 / 1.20 / 0.06 / 0.375; Muse Spark 1.3 Contributor, Muse Spark 1.2 Contributor 0.10 / 0.20 / 0.002; Qwen3.8 Max 2.00 / 6.00 / 0.25 / 2.50; Qwen3.8 Flash 0.15 / 0.47 / 0.016 / 0.20; Qwen3.7 Plus (≤256K tokens) 0.40 / 1.60 / 0.04 / 0.50, (>256K tokens) 1.20 / 4.80 / 0.12 / 1.50; DeepSeek V4.1 Flash (off-peak) 0.15 / 0.60 / 0.003, (peak) 0.30 / 1.20 / 0.006; DeepSeek V4 Pro (off-peak) 0.66 / 1.98 / 0.022, (peak) 1.32 / 3.96 / 0.044 (peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday); Hy4 preview 0.834 / 2.501 / 0.042; Hy3 0.14 / 0.58 / 0.035; Space Bunny 0.15 / 0.60 / 0.03; Grok 4.6, Grok 4.7 (≤200K tokens) 2.00 / 6.00 / 0.50, (>200K tokens) 4.00 / 12.00 / 1.00; GPT 6 Luna (≤272K tokens) 0.10 / 0.50 / 0.01 / 0.125, (>272K tokens) 0.20 / 0.75 / 0.02 / 0.25; GPT 5.6 Luna (≤272K tokens) 0.20 / 1.20 / 0.02 / 0.25, (>272K tokens) 0.40 / 1.80 / 0.04 / 0.50; Claude Haiku 5.5 (≤100K tokens) 0.10 / 0.50 / 0.01 / 0.125, (>100K tokens) 0.50 / 2.50 / 0.05 / 0.625; Gemini 4 Argon Contributor 4.00 / 20.00 / 0.20
  • Client requirement: requests must carry the session header x-opencode-session for routing and prompt-cache optimisation; the docs list validated clients (Hermes, Claude Code, Codex, ZCode, Pi, jcode, Kilo Code CLI) and known problematic clients (DeepSeek Harness, GitHub Copilot Chat, Kimi Code, MiMo Code)
  • Supported models: Grok 4.7, Grok 4.6, GLM-5.3/5.3-Flash/5.2, GPT 6 Luna, GPT 5.6 Luna, Claude Haiku 5.5, Gemini 4 Argon Contributor, Kimi K3/K2.7 Code/K2.6, LongCat-2.0, MiMo-V2.6-Flash/V2.6-Pro/V2.5/V2.5-Pro, MiniMax M3/M2.7, Muse Spark 1.3 Contributor, Muse Spark 1.2 Contributor, Qwen3.8 Max/Qwen3.8 Flash/Qwen3.7 Plus, DeepSeek V4.1 Flash/V4 Pro, Hy4 preview, Hy3, Space Bunny, LongCat 2.5 Preview Free (limited-time free), Step 5 Preview Free (limited-time free), Cat Jam Free (limited-time free)
  • Two models added on 2026-10-08: Claude Haiku 5.5 — tiered at 100K tokens, ≤100K tokens input $0.10 / output $0.50 / cache read $0.01 / cache write $0.125 and >100K tokens input $0.50 / output $2.50 / cache read $0.05 / cache write $0.625 per 1M tokens, monthly usage allowance $15 on Go and $60 on Go Plus, estimated request limits 3,850 / 5 hours, 9,620 / week and 19,230 / month on Go and 15,380 / 38,460 / 76,920 on Go Plus, model ID claude-haiku-5-5, endpoint https://opencode.ai/zen/go/v1/messages (@ai-sdk/anthropic); Step 5 Preview Free — input/output/cache read all free with unlimited requests and monthly allowance (limited-time), model ID step-5-preview-free, endpoint https://opencode.ai/zen/go/v1/chat/completions
  • 2026-10-14 model-library change (33 → 32 models): DeepSeek V4 Flash and DeepSeek V4 Flash Vision Exp were removed (their rows disappear from the token-price table, both monthly-allowance tables, the estimated-request table, the endpoint table and the privacy table, their two lines are dropped from the per-request token estimates, and the DeepSeek peak/off-peak note changes from "DeepSeek V4.1 Flash / V4 Pro / V4 Flash / V4 Flash Vision Exp" to "DeepSeek V4.1 Flash / V4 Pro"), and Gemini 4 Argon Contributor was added — token prices input $4.00 / output $20.00 / cache read \(0.20 per 1M tokens (no cache-write price), monthly allowance Go ~~\)30~~ \(60** and Go Plus ~~\)120~~ $240** (both marked "2x · Limited time", i.e. a limited-time doubling), estimated requests 440 per 5 hours, 1,080 per week and 2,120 per month on Go and 1,760 / 4,320 / 8,480 on Go Plus, model ID gemini-4-argon-contributor, endpoint https://opencode.ai/inference/go/v1/models/gemini-4-argon-contributor (@ai-sdk/google, the only model in the doc using the /v1/models/<model-id> form), and a privacy row of Model training Yes with Data retention Not ZDR plus the note "Your prompts and completions may be used to train future Google models." No other model's price, allowance or request limit changed, and the two crawled URLs (/docs/zh-cn/go and /v2/docs/console/go/) serve identical content
• StepFun International Coding Plan
• UniAI GLM-5 Coding Plan
• Moorethreads AI Coding Plan
• KwaiKAT Coding Plan
• DeepSeek API Pricing:
  • Peak-valley pricing: off-peak prices are half of peak; peak hours are Monday to Friday 9:00-12:00 and 14:00-18:00 Beijing time
  • deepseek-flash (DeepSeek-V4.1-Flash, 1M context, supports image understanding): off-peak cached input 0.02 / uncached input 1 / output 4 RMB per 1M tokens; peak 0.04 / 2 / 8
  • deepseek-v4-pro (DeepSeek-V4-Pro-0813, 1M context, no image understanding): off-peak cached input 0.15 / uncached input 4.5 / output 13.5 RMB per 1M tokens; peak 0.30 / 9.0 / 27.0

Prompts, Requests, and Tokens

• Prompt: User inputs a prompt into the CLI and presses Enter to send it. From the request perspective, it's when the last message is from the user, not from a tool call result.
• Request: In addition to the prompt itself generating one request, after each round of tool calls, the tool call results are sent along with the context until there are no more tool calls.
• Token: Each request has a certain amount of input and output tokens.

One prompt corresponds to multiple requests, and each request has many input and output tokens. Some input tokens will hit the cache. In actual testing in Vibe Coding scenarios, among input + output tokens:

• Input tokens account for 99.5%, because with multi-turn conversations, input tokens accumulate and are repeatedly counted
  • Among which cached tokens account for approximately 90-95% of input + output tokens
• Output tokens account for 0.5%

Common API Pricing Models

• OpenAI Model: Automatic caching, with uncached input price, cached input price, and output price
  • OpenAI has three prices: Input, Cached Input, and Output. If the access doesn't hit the cache, the uncached portion is charged at the Input rate; OpenAI may perform caching; if the access hits the cache, the cached portion is charged at the Cached Input rate.
  • Typically Cached Input is 0.1x the Input price, sometimes between 0.1-0.2x.
• Anthropic Model: Manual caching, with uncached input price, cached input price, input price with cache write (different TTLs may correspond to different prices), and output price
  • Claude has five prices: Base Input Tokens, 5m Cache Writes, 1h Cache Writes, Cache Hits & Refreshes, and Output Tokens. If caching is not used, each input is charged at the Base Input Tokens rate; if caching is used, the portion written to cache is charged at the 5m/1h Cache Writes rate, and subsequent cache hits are charged at the Cache Hits & Refreshes rate.
  • Currently 5m Cache Writes are 1.25x the Base Input Tokens price, 1h Cache Writes are 2x the Base Input Tokens price, and Cache Hits & Refreshes are 0.1x the Base Input Tokens price.

Model Parameters Comparison

Name	Parameters	Active	Vision
DeepSeek-V4-Flash-0731	284B	13B	N
DeepSeek-V4-Flash	284B	13B	N
DeepSeek-V4-Pro-0813	1.6T	49B	N
DeepSeek-V4-Pro	1.6T	49B	N
DeepSeek-V4.1-Flash	552B	8B input / 16B output	Y
GLM-4.7-Flash	30B	3B	N
GLM-4.7	355B	32B	N
GLM-5.1	744B	40B	N
GLM-5.2	744B	40B	N
GLM-5.3-Flash	320B	18B	Y
GLM-5.3	744B	40B	N
Hy3-preview	295B	21B	N
Kimi-K3	2.8T	104B	Y
Kimi-K2.6	1T	32B	Y
MiniMax-M3	428B	23B	Y
MiniMax-M2.7	230B	10B	N
Qwen3.8-2.4T-A95B	2.4T	95B	Y
Qwen3.8-27B	27B	-	Y
Qwen3.8-Flash-Next	180B	6B	Y
Qwen3.5-397B-A17B	397B	17B	Y

Update History

• 2026/10/14: OpenCode Go changed its model library (33 → 32 models): DeepSeek V4 Flash and DeepSeek V4 Flash Vision Exp were removed and Gemini 4 Argon Contributor was added. The new model's token prices are input $4.00 / output $20.00 / cache read \(0.20 per 1M tokens (no cache-write price), its monthly allowance is Go ~~\)30~~ \(60** and Go Plus ~~\)120~~ $240** (both marked "2x · Limited time", i.e. a limited-time doubling), its estimated requests rise from 220 to 440 per 5 hours, 540 to 1,080 per week and 1,060 to 2,120 per month on Go and 880 → 1,760, 2,160 → 4,320, 4,240 → 8,480 on Go Plus, model ID gemini-4-argon-contributor, endpoint https://opencode.ai/inference/go/v1/models/gemini-4-argon-contributor (@ai-sdk/google, the only model in the doc using the /v1/models/<model-id> form), and a privacy row of Model training Yes with Data retention Not ZDR plus the note "Your prompts and completions may be used to train future Google models." The two removed DeepSeek models disappear from the supported-model list, the token-price table, both monthly-allowance tables, the estimated-request table, the endpoint table and the privacy table, their two lines are dropped from the per-request token estimates, and the DeepSeek peak/off-peak note changes from "DeepSeek V4.1 Flash / V4 Pro / V4 Flash / V4 Flash Vision Exp" to "DeepSeek V4.1 Flash / V4 Pro"; no other model's price, allowance or request limit changed, and the two crawled URLs (/docs/zh-cn/go and /v2/docs/console/go/) serve identical content

• 2026/10/14: MiniMax completed the M Plan migration of its subscription-pricing and pay-as-you-go pages (four pages across the Chinese and English sites). /docs/guides/pricing-token-plan (subscription pricing) was rewritten from the old Token Plan content (Plus ¥49 / Max ¥119 / Ultra ¥469 per month, internationally $22 / $55 / $132; credit packs ¥30 → 4,489, ¥150 → 22,460 and ¥500 → 74,900 credits, internationally $5 / $25 / $100, valid for 365 days) into an M Plan overview page with the same structure as /docs/m-plan/intro — an available-models table splitting the Go / Explore / Build tiers (the M3.1 Flash Preview text model on all three, the H3 video model on Explore and Build only, image and audio models on all three), a usage table of Go / Explore 3x / Build (12x on the Chinese site, 7.5x internationally), and supported products listing MiniMax Code (officially recommended: sign in to get plan benefits with no separate subscription or API Key) plus Claude Code, Codex, Cursor, OpenCode, Pi, Hermes Agent, OpenClaw and other tools — and the price table and the credit-pack table have been removed from that page. On both sites the pay-as-you-go page /docs/guides/pricing-paygo changed from "pay as you go / API price" to "pricing overview / product pricing", with the site navigation regrouped into model pricing (language, audio, video, image, music), tool pricing (MCP, server-side tools), packages (audio packages, video packages) and billing guides (payment-method comparison, pay-as-you-go billing, package usage rules, FAQ); the body gained the audio packages (HD series ¥630 / ¥5,950 / ¥56,000 and Turbo series ¥360 / ¥3,400 / ¥32,000, valid for 1 month / 3 months / 1 year), the video packages, MCP and server-side tools and the billing guides — sections that had not been crawled before — and every "Token Plan" mention in the body became "M Plan" (e.g. the M Plan MCP plugin note for API-vlm). Every existing price figure on the two pay-as-you-go pages is unchanged (MiniMax-M3 ≤512K input tokens permanently 50% off at 2.10 / 8.40 / 0.42 and >512K at 4.20 / 16.80 / 0.84, with cache reads discounted to the same degree), and the plan tiers, usage rules and M3 API pricing are unchanged

• 2026/10/14: The Alibaba Cloud Bailian Token Plan (Personal Edition) pages (Token Plan overview, Personal Edition overview and Personal Edition FAQ, all three last updated 2026-10-14 16:03:59) reworded the monthly quota from "a subscription period running 30 days forward from the subscription date" to "the quota is counted by subscription month", and in the FAQ "each 30 days starting from the subscription date counts as one subscription month" became "the quota auto-refreshes by subscription month", with the worked example of a Standard plan subscribed on 10 September (that subscription month running 10 September to 10 October) deleted; the FAQ footer's "previous" link also changed from "Connecting to the Harness tool" to "Harness benefits" (/zh/model-studio/token-plan-harness-tool → /zh/model-studio/token-plan-harness-benefits). The rule itself is unchanged — still metered by subscription month, still resetting based on the subscription date rather than a fixed calendar date, still pausing service once the quota is exhausted with unused quota not rolling over (the official docs no longer state the specific "30 days"); the tiers and prices (Lite 60 / Essential 120 / Standard 180 / Pro 600 RMB per month, 39/79/139/499 for a limited time), the per-tier Credits allowances, the usage pack, the deduction order and upgrade rules, the supported models and the limited-time night discounts are all unchanged on the three pages

• 2026/10/14: Volcano Engine extended both deduction-promotion windows (all three pages last updated 2026-10-14 10:12:38–10:12:39): the deepseek-v4.1-flash 50%-off deduction promotion deadline moved from 2026-10-30 18:00 to 2026-12-30 23:59:59 (start unchanged — Agent Plan 2026-09-15 00:00, Coding Plan 2026-09-23 00:00; coefficient 2.5 → 1.25), and the kimi-k2.8-preview 40%-off deduction promotion (Agent Plan, coefficient 8 → 4.8) plus the Coding Plan campaign during which that model's available quota is "comparable to that during the Agent Plan 40%-off deduction promotion" both moved from 2026-10-14 23:59 to 2026-12-14 23:59:59 (starts unchanged — Agent Plan 2026-09-17 00:00, Coding Plan 2026-09-18 00:00). The Agent Plan (Personal Edition) overview, the Agent Plan (Personal Edition) billing page and the Coding Plan (Personal Edition) overview were updated together (the old 10-30 / 10-14 deadlines expired today, so both campaigns were about to end); the kimi-k3 50%-off campaign (2026-10-12 00:00 to 2026-10-25 23:59), every other model's deduction coefficient, the plan tiers and prices, the model library and the 1M-context support list are all unchanged

• 2026/10/14: OpenCode Go updated the data policy of its limited-time free model Cat Jam Free: in the docs privacy table its "Model training" column changed from Yes to Not used and its "Data retention" column from Not ZDR to 30 days, and the corresponding note "Cat Jam Free: During its free period, collected data may be used to improve the model." was removed, leaving only the "Cat Jam Free: Free for a limited time." label. The model remains in the supported-model list, the token-price table (input/output/cache read all Free) and the request/monthly-allowance tables (both unlimited), with the same model ID cat-jam-free and endpoint; the other models' privacy rows, prices and allowances are unchanged (both crawled URLs /docs/zh-cn/go and /v2/docs/console/go/ are in sync)

• 2026/10/14: OpenCode Go added the limited-time free Cat Jam Free model: input, output and cache read are all listed as Free, and both the request limits and the monthly usage allowance are labelled unlimited (a limited-time promotion with no end date given by the vendor); model ID cat-jam-free, endpoint https://opencode.ai/inference/go/v1/chat/completions. The supported-model list (32 → 33 models), the token-price table, both monthly-allowance tables, the estimated-request table, the endpoint table and the privacy table all gained this model; in the privacy table its "Model training" column is Yes and its "Data retention" column is Not ZDR, and the page adds the note "Cat Jam Free: During its free period, collected data may be used to improve the model." No other model's price, allowance or request limit changed, and the two crawled URLs (/docs/zh-cn/go and /v2/docs/console/go/) serve identical content (differing only in the source-URL comment)

• 2026/10/14: Volcano Engine marked kimi-k2.7-code as "being retired", synchronised across three Coding Plan / Agent Plan pages (pages last updated 2026-10-13 21:57): the model description table on the Coding Plan (Personal Edition) overview, the model table on the Agent Plan (Personal Edition) overview, and the AFP deduction-coefficient table in the Agent Plan billing page all now append "being retired" to that model name (its deduction coefficient remains 4.5 and all tiers remain ticked; no retirement date has been published yet). The Coding Plan (Personal Edition) announcements/changelog page added a 10.13 entry, "model deprecation started: kimi-k2.7-code", linking to the model-deprecation page. The three plan pages and the changelog page all carry 最近更新时间 = 2026-10-13 21:57:48–21:57:50; plan tiers and prices, all other models and their deduction coefficients, and the 1M-context support list are unchanged. The accompanying integration doc (connecting Coding Plan to OpenClaw) swapped the model in its image-input configuration example from kimi-k2.7-code to doubao-seed-2.1-lite — an example-only change with no rule change

• 2026/10/13: The Chinese-site GLM Coding Plan overview page (docs.bigmodel.cn/cn/coding-plan/overview and its .md variant) added a plan upgrade and renewal clause to its point-refresh-rules section: after a subscriber upgrades, the points allowance refreshes immediately at the new tier, while renewal does not trigger a points reset and the existing refresh cycle is unchanged. Nothing else on the page changed: plan tiers and prices (Lite 118 / Pro 538 / Max 1078 RMB per month), the 5-hour and weekly points allowances (2000/10000, 12000/60000, 28000/140000), the model and MCP deduction coefficients, the supported models (GLM-5.3, GLM-5.3-Flash), the usage reference table, the off-peak 50% deduction rule, the compatible-tool list and the "save up to 92%" claim are all unchanged. This is a rules addition covering how allowances refresh, and since the KB's main entry did not previously record Zhipu's point-refresh information, that section (the dynamic 5-hour refresh, the weekly cycle and the new upgrade/renewal rule) was added to it in this round

• 2026/10/13: The OpenCode Go docs page moved to v2 (https://opencode.ai/v2/docs/console/go/) — the old URL /docs/zh-cn/go (which returned 404 on 2026-10-12) now 302-redirects there, so the Go content moves back from the marketing page (/zh/go, which lists only 13 featured models) to the docs site. With the page restored, the docs publish the full data set for the first time: token-price, monthly-allowance, estimated-request (5-hour/week/month), per-request token estimate and data-retention tables for 32 models. Changes versus the KB's previous data: the endpoint domain changed from opencode.ai/zen/go/v1/… to https://opencode.ai/inference/go/v1/… (the two domains are aliases; /v1/models returns 200 on both), and the docs now document referencing models as opencode-go/ inside OpenCode plus the model-list endpoint https://opencode.ai/inference/go/v1/models; apart from DeepSeek V4.1 Flash's \(80/\)160 (raised on 2026-10-12), the monthly allowances fall into 6 tiers — \(15/\)60, \(30/\)90, \(30/\)120, \(60/\)120, \(60/\)180 and \(60/\)240 — and the supported-model list is unchanged. The page also adds a Validated clients table (Hermes, Claude Code, Codex, ZCode, Pi, jcode, Kilo Code CLI) and a Known problematic clients table (DeepSeek Harness, GitHub Copilot Chat, Kimi Code, MiMo Code), both about x-opencode-session header support
• 2026/10/13: The MiniMax M Plan intro pages (domestic and international /docs/token-plan/intro) were restructured, adding Pi to the compatible-tool list: the "MiniMax Code · Official recommendation" item was promoted from a bullet in the tool list to its own section (with the copy "use MiniMax's official agent app for local development and everyday work; sign in to access your M Plan benefits, with no separate subscription or API Key setup"), and the tool list changed from Claude Code, Codex, Cursor, OpenCode, Hermes Agent, OpenClaw to the same list with Pi inserted between OpenCode and Hermes Agent; plan tiers and prices, usage rules, credit packs and migration rules are unchanged
• 2026/10/13: Volcano Engine added a limited-time 50%-off deduction coefficient for kimi-k3: from 2026-10-12 00:00 to 2026-10-25 23:59 its Agent Plan deduction coefficient is halved, i.e. 10 → 5. The Agent Plan (Personal Edition) overview (campaign announcement) and billing pages (the AFP coefficient table now reads "10 (5 at a limited-time discount)") were updated together; the other Agent Plan models' deduction coefficients, the plan tiers and prices, and the Coding Plan (Personal Edition) pages are unchanged (the Coding Plan page did not carry this campaign, so its promotions are still only the deepseek-v4.1-flash 50% off and the Kimi-K2.8-Preview quota campaign)
• 2026/10/13: Volcano Engine removed three models previously marked "being retired" from its model lists: doubao-seed-2.1-turbo and doubao-seed-2.0-lite (both previously listed in the Coding Plan (Personal Edition) supported-model table, the Agent Plan (Personal Edition) model table and the AFP deduction-coefficient table; their rows are now deleted, and the doubao-seed-2.1-turbo reference in the Agent Plan billing text-deduction example was changed to doubao-seed-evolving) and doubao-seedream-5.0-lite (removed from the Agent Plan image-generation coefficient table along with its image-generation deduction example, leaving doubao-seedream-5-0-pro as the only image model). The three pages involved — the Coding Plan Personal overview, the Agent Plan Personal overview and the Agent Plan Personal billing doc — share the same last-updated timestamp of 2026-10-12 21:22:20; plan tiers and prices, all other models and their deduction coefficients, and the 1M-context support list are unchanged
• 2026/10/13: MiniMax published a retirement notice for the M2.5 and M2.7 model families on its models overview pages (platform.minimaxi.com and platform.minimax.io, /docs/guides/models-intro): from 2026-10-19 the models are no longer available to new users, and on 2026-11-18 they are retired. Existing pay-as-you-go API users may keep calling them until the retirement date, after which API requests to these models no longer work; for M Plan and Token Plan users, requests to M2.5 and M2.7 are routed to M3.1 Flash Preview from 2026-10-19, with the original model names kept as compatible entry points (no model-name changes needed) and plan usage entitlements unaffected. The two sites carry identical notices (differing only in language), and the M Plan tiers' supported models, prices and allowances are unchanged
• 2026/10/12: Alibaba Cloud Bailian Token Plan (Personal Edition) added wan3.0-video to its supported models (Wanxiang brand, capability listed as "video generation", placed after wan2.7-image and wan2.7-image-pro; overview page updated 2026-10-12 18:20:44). Plan tiers and prices (Lite 60 / Essential 120 / Standard 180 / Pro 600 RMB per month, 39/79/139/499 for a limited time), the per-tier Credits allowances, the usage pack, the deduction order and upgrade rules, the limited-time night discounts (40% off qwen3.8-max/qwen3.8-flash and 50% off deepseek-v4-pro-0813/deepseek-v4-flash-0731/deepseek-v4.1-flash) and the limited-time free decision-model-preview are all unchanged; the Team Edition overview page did not change in this round and still does not list wan3.0-video
• 2026/10/12: OpenCode Go raised the usage allowance for DeepSeek V4.1 Flash, and a site revamp moved the Go/Zen pages out of the docs site. The former docs page https://opencode.ai/docs/zh-cn/go now returns 404 (neither the docs sitemap nor the docs sidebar contains go/zen routes any more; the site's own navigation now points at /go and /zen, with the content hosted on marketing pages), and the new page is https://opencode.ai/zh/go — which still keeps a link to the dead /docs/zh-cn/go#pricing; the link in the KB was updated accordingly. The new page's "usage" table lists the estimated requests per 5 hours and the monthly usage allowance (Go / Go Plus) for 13 featured models: DeepSeek V4.1 Flash went from $60 / $120 to $80 / $160, with its estimated 5-hour requests rising from 26,000 / 52,000 to 34,670 / 69,330 — both the allowance and the two tiers' request figures are exactly 4/3 of the old values, so the estimated per-request usage is unchanged. The other 12 models match the old docs page exactly: Kimi K3 $15 / $60, Kimi K2.7 Code $60 / $180, Space Bunny $30 / $120, MiniMax M3 $60 / $180, Claude Haiku 5.5 $15 / $60, GPT 6 Luna $15 / $60, Qwen3.7 Plus $60 / $180, GLM-5.3-Flash $60 / $180, MiMo-V2.6-Flash $60 / $120, Muse Spark 1.3 Contributor $60 / $120, and LongCat 2.5 Preview Free and Step 5 Preview Free with no limit. Both language versions (/zh/go and /go) carry identical figures. The new page does not publish per-model token prices, and the old docs page's token-price and weekly/monthly request tables could not be re-checked because that page is gone, so whether DeepSeek V4.1 Flash's token prices changed in step, and whether any model not featured on the new page changed, remain unknown (the corresponding KB data still comes from the last successful crawl at 12:30 today)
• 2026/10/11: The Alibaba Cloud Bailian model pricing page (updated from 2026-10-10 17:27:17 to 2026-10-11 11:09:03) kept pruning legacy model entries, with the unit prices and billing rules of the remaining models unchanged: it dropped qwen-plus-2025-09-11, qwen-plus-2025-07-28, qwen3-omni-flash-2025-09-15, qwen3-omni-flash-realtime (the undated row, keeping qwen3-omni-flash-realtime-2025-09-15), qwen3-vl-flash-2026-01-22 (US), qwen3-vl-plus-2025-09-23 (Global and EU) along with the "current capabilities equivalent to qwen3-vl-plus-2025-12-19" note on the qwen3-vl-plus row (the row itself stays), qwen3-asr-flash-2025-09-08, fun-asr-2025-08-25, qwen3-livetranslate-flash-realtime and its 2025-09-22 version, plus the last two deepseek-v4-flash rows (Global and Japan) in the DeepSeek section; the "Text generation — Qwen open source" chapter dropped four whole sections — Qwen3 (qwen3-next-80b-a3b-thinking / -instruct, qwen3-235b-a22b-thinking-2507, qwen3-30b-a3b-thinking-2507 / -instruct-2507, qwen3-235b-a22b, qwen3-30b-a3b, qwen3-14b, qwen3-8b), Qwen3-Omni-Captioner (qwen3-omni-30b-a3b-captioner), Qwen-VL (qwen3-vl-235b-a22b-thinking, qwen3-vl-32b-thinking / -instruct, qwen3-vl-30b-a3b-thinking / -instruct, qwen3-vl-8b-thinking / -instruct) and Qwen-Coder (qwen3-coder-480b-a35b-instruct, qwen3-coder-30b-a3b-instruct) — so the open-source chapter now only keeps Qwen3.8, Qwen3.6, Qwen3.5 and Qwen-Audio; the Qwen-Omni free-quota note changed from "the Qwen3.5-Omni, Qwen3-Omni and Qwen-Omni-Turbo series" to "the Qwen3.5-Omni and Qwen3-Omni series". Everything dropped is a legacy version or legacy open-source model in Bailian's general model catalogue, and none of it was in the Token Plan Personal or Team Edition supported-model lists (deepseek-v4-flash-0731 remains in both); no coding plan page changed on the same day, so there is no user-visible impact on coding plans
• 2026/10/11: The z.ai international release-notes page (docs.z.ai/release-notes/new-released — its first change since 2026-08-28) was tidied up, dropping a batch of older model entries: GLM-OCR, GLM-4.7-Flash, AutoGLM-Phone-Multilingual, GLM-ASR-2512, GLM-4.6V, GLM-4.5V, GLM Slide/Poster Agent (beta), the GLM-4.5 series and CogVideoX-3, leaving GLM-5.3-Flash, GLM-5.3, GLM-5.2, GLM-5.1, GLM-5, GLM-Image, GLM-4.7 and GLM-4.6. The Chinese-site counterpart (docs.bigmodel.cn/cn/update/new-releases) still lists every one of those entries, and neither the international DevPack overview page nor the bigmodel.cn pricing page changed in this round, so this is a release-notes page tidy-up rather than a model delisting and it alters no coding plan's supported models, with no user-visible impact
• 2026/10/11: The Chinese-site GLM Coding Plan overview page (docs.bigmodel.cn/cn/coding-plan/overview and its .md variant) changed its positioning copy from "only a small outlay covers the whole development flow of requirement understanding, code generation, debugging and fixing, codebase Q&A and automated task handling" to "only a small outlay covers the whole flow from code generation and debugging/fixing to autonomous task decomposition and delivery — Coding and Work in one plan", i.e. the claimed scope was widened from a coding workflow to "Coding and Work in one plan" (in line with the international site's 2026-10-09 change from "coding tools" to "agent tools" and its new Office Work section). Everything else on the page is unchanged: plan tiers and prices (Lite 118 / Pro 538 / Max 1078 RMB per month), the 5-hour and weekly points allowances, the model and MCP deduction coefficients, the supported models (GLM-5.3, GLM-5.3-Flash, with GLM-5.2/5.1 auto-switching to GLM-5.3 and GLM-5-Turbo/GLM-4.7 to GLM-5.3-Flash), the usage reference table, the compatible-tool list (ZCode, AutoClaw, Claude Code, Codex, OpenCode, WorkBuddy) and the "save up to 92%" claim, so there is no user-visible impact
• 2026/10/10: Alibaba Cloud Bailian delisted a batch of older models. The model pricing page (updated 2026-10-10 17:27:17) dropped a large number of legacy entries with the unit prices and billing rules of the remaining models unchanged: the qwen3-max family (qwen3-max, qwen3-max-2026-01-23, qwen3-max-2025-09-23, qwen3-max-preview), qwen3.6-max-preview, qwen3-coder-plus, qwen3-coder-next, qwen-turbo, qwq-plus, qvq-max/qvq-plus, qwen-math-plus/qwen-math-turbo, qwen-long-latest/qwen-long-2025-01-25, the old Qwen-Omni series (qwen-omni-turbo and its versions, qwen3-omni-flash-2025-12-01/2025-09-15, the qwen-omni-turbo-realtime series), qwen3-vl-flash, qwen-vl-max, qwen-vl-plus, deepseek-v4-flash (the China-site and "Global"/"US"/"International" rows; only the "Global" and "Japan" deployment rows remain), deepseek-v3.2, deepseek-v3.2-exp, deepseek-v3.1, deepseek-v3, deepseek-r1, deepseek-r1-distill-qwen-7b/14b/32b and llama-8b (already marked discontinued), glm-4.7, glm-4.6, glm-4.5, glm-4.5-air, MiniMax-M2.1, kimi-k2-thinking, Moonshot-Kimi-K2-Instruct, the old qwen-image-max/qwen-image-plus and qwen-image-edit series plus legacy image sections such as "image background generation" and "image outpainting", and the China-site rows for open-source models such as qwen3-32b and qwen3-vl-32b/30b/8b; the context-cache page (updated 2026-10-10 15:08:27) likewise removed deepseek-v3.2, deepseek-v4-flash, qwen3-coder-plus, qwen-turbo, qwen3-vl-flash, qwen-vl-max/qwen-vl-plus, glm-4.6 and MiniMax-M2.1 from its supported-model lists. Impact on coding plans: the integration docs (OpenCode and OpenClaw, updated 2026-10-10 15:08:29) removed deepseek-v4-flash and deepseek-v3.2 from their bailian-token-plan/* model lists; the Token Plan Team Edition overview page (/zh/model-studio/token-plan-team-overview, not part of the scheduled crawl — fetched manually for this check) no longer lists either model and does include glm-5.3 and deepseek-v4.1-flash (the main entry was updated to the page's current content); the Token Plan Personal Edition supported-models page is unchanged in this round, as it already listed only deepseek-v4-flash-0731 and never deepseek-v4-flash or deepseek-v3.2
• 2026/10/10: The Alibaba Cloud Bailian Coding Plan cut back the Pro plan's supported models (plan page updated 2026-10-10 15:08:26): the official docs dropped qwen3-max-2026-01-23, qwen3-coder-next, qwen3-coder-plus and glm-4.7 from the "more models" tier, which now lists qwen3.5-plus (image understanding) alone; the recommended models (qwen3.7-plus, qwen3.6-plus, kimi-k2.5, glm-5, MiniMax-M2.5), the price (¥200/month, ¥39.90 for the first month for new subscribers) and the request limits (6,000 per 5 hours, 45,000 per week, 90,000 per month) are unchanged. The rest of the docs were updated in step: the FAQ page (also updated 15:08:26) removed those four models from its context-length table (qwen3-max-2026-01-23 262,144, qwen3-coder-next 262,144, qwen3-coder-plus 1,000,000, glm-4.7 202,752) and dropped qwen3-coder-plus from its "switch model" advice, and the OpenClaw configuration example's model ID changed from qwen3-max-2026-01-23 to qwen3.7-max; the integration docs (OpenCode and OpenClaw, updated 2026-10-10 15:08:29) removed the same four models from their bailian-coding-plan/* list, which now contains qwen3.7-plus, qwen3.6-plus, qwen3.5-plus, MiniMax-M2.5, glm-5 and kimi-k2.5
• 2026/10/10: The CTCloud Programming Token Plan's previous token-metered plans (renewable by existing subscribers only) completed the delisting that had been scheduled for that day: GLM-5.0 and DeepSeek-V3.2 (written "DeepSeeV3.2" on the page), previously marked as "retiring 10 October", were dropped from the list while GLM-5.2 was added, leaving DeepSeek-V4-Flash-0731, GLM-5.1 and GLM-5.2. The Credits Edition's tiers and prices (29/89/199/399/699 RMB per month), credit allowances, supported models (DeepSeek-V4-Pro, DeepSeek-V4-Flash-0731, GLM-5.2, GLM-5.1, Kimi-K2.6, MiniMax-M3) and credit-conversion table are unchanged
• 2026/10/09: Alibaba Cloud Bailian added the image model qwen-image-2.1-turbo: 0.1 CNY per image on the China site and the "Global" deployment (10 free images) and 0.113362 CNY per image internationally, listed alongside qwen-image-2.1-pro (0.25 / 0.283404 CNY per image); the unit prices of the rest of the qwen-image-2.0 series and of all text models are unchanged. This change only concerns image-generation pricing in Bailian's general model catalogue and alters no coding plan's model library, tiers or quotas, the Token Plan included
• 2026/10/09: The Kimi Code API gained support for the OpenAI Responses protocol: the OpenAI-compatible interface now offers both the Chat Completions and Responses protocols, with the endpoint https://api.kimi.com/coding/v1/responses on the domestic domain (no overseas example given). The Base URLs, the Anthropic-compatible endpoints, model IDs and API prices are unchanged
• 2026/10/09: Infini (Wuwen Xinqiong) GenStudio took deepseek-v4-flash offline at 12:00 on 2026-10-09 as previously announced, with the recommended replacement model still deepseek-v4.1-flash. The platform's Infini Coding Plan was discontinued on 2026-06-27, so this change only affects GenStudio API calls
• 2026/10/09: Zhipu GLM Coding Plan's two limited-time campaigns both expired: the "GLM-5.3-Flash Usage Campaign" (2026-09-03 to 2026-10-07, unlimited GLM-5.3-Flash usage on ZCode/AutoClaw with doubled quota on other agents) and the "Double Festival" campaign (2026-09-25 to 2026-10-07, all-day usage billed at the off-peak rate). Both sites are back to the regular rule (off-peak hours consume points at 50% of the base rate; peak hours are weekdays 14:00–18:00 (UTC+8)), with no campaign running. The Chinese site changed its copy from "a subscription plan built for AI coding" to "a subscription plan for large-model usage", and rewrote the compatible-tool list from "Claude Code, Kilo Code, OpenClaw, OpenCode, TRAE, CodeBuddy" to "ZCode, AutoClaw, Claude Code, Codex, OpenCode, WorkBuddy". The international version changed "coding tools" to "agent tools", with the applicable tools changed from "Claude Code, Cline, OpenCode" to "ZCode, Claude Code, Codex, OpenCode, AutoClaw, OpenClaw", added an "Office Work" section (Deep Search & Research, Business Data Analysis, Creative Ideation & Brainstorming, Automated Task Execution), changed the two "Advantages" headings to "Access to High-Intelligence Models" and "Works with 20+ Agent Tools", and changed the FAQ quota-sharing wording from "All supported coding tools share the same usage quota" to "All supported tools share the same usage quota". Plan tiers and prices (from 18 USD per month), points allowances, deduction coefficients, model coverage and the "save up to 92%" claim are unchanged
• 2026/10/08: The Token Plan intro and FAQ pages on both the Chinese and English MiniMax sites (/docs/token-plan/intro and /docs/token-plan/faq on platform.minimaxi.com and platform.minimax.io — four pages) were rewritten from the old Token Plan content to M Plan content, effectively completing the site migration. The pages state that the Build tier's allowance is 12x Go on the Chinese site while the English site still says 7.5x (the two sites disagree; the main entry above records them separately). The tier model coverage (all three tiers include the M3.1 Flash Preview text model plus image/audio generation, web search and multimodal understanding; Explore and Build add the H3 video model while Go has no video) and the rules "non-video models are subject to 5-hour and weekly windows, video models only to the weekly window" and "1,000 credits = ¥7, valid for 1 year" are unchanged. Points added by the FAQ include: the Subscription Key works with any OpenAI- or Anthropic-compatible tool (Claude Code, Codex, Cursor, OpenCode, Hermes Agent and OpenClaw are listed), MiniMax Code needs no separate subscription or Subscription Key, downgrades are not supported (monthly plans upgrade to a higher monthly or annual tier; annual plans only to a higher annual tier), auto-renewal cannot be restored once disabled, both the 5-hour and weekly windows start on first use and refresh independently, exhausting the quota never charges the account balance automatically, and a 401 after login can be fixed by setting the region manually with mmx config set --key region. Meanwhile /docs/guides/pricing-token-plan (subscription pricing) and /docs/guides/pricing-paygo are still on the old Token Plan / pay-as-you-go content and have not been migrated yet
• 2026/10/08: Baidu Qianfan Coding Plan cut its model library down sharply and announced that model updates have stopped: the deduction-coefficient table and the configurable model list now contain only DeepSeek-V4-Pro (peak 5 / off-peak 4), GLM-5.1 (peak 4 / off-peak 3) and Baidu Search (10), removing Kimi-K2.5, DeepSeek-V3.2, DeepSeek-V4-Flash, GLM-5, MiniMax-M2.5 and ERNIE-4.5-Turbo-20260402 (the Model Names that can be set in the config file dropped from six to deepseek-v4-pro and glm-5.1); the previous "DeepSeek-V4-Flash and GLM-5.1 are trial versions with limited resources" sentence was replaced by the new notice "the supported model range will be adjusted dynamically with model iterations. Coding Plan has stopped model updates; users who need newer models are advised to migrate to Token Plan (Personal Edition)". Plan prices and limits (Lite ¥40, Pro ¥200 per month, 1,200 / 6,000 requests per 5 hours) are unchanged
• 2026/10/08: OpenCode Go added Claude Haiku 5.5 and the limited-time free Step 5 Preview Free. Claude Haiku 5.5 is tiered at 100K tokens (≤100K tokens input $0.10, output $0.50, cache read $0.01, cache write $0.125; >100K tokens input $0.50, output $2.50, cache read $0.05, cache write $0.625 per 1M tokens), with a monthly usage allowance of $15 on Go and $60 on Go Plus, estimated request limits of 3,850 / 5 hours, 9,620 / week and 19,230 / month on Go and 15,380 / 38,460 / 76,920 on Go Plus, a per-request usage estimate of 1,100 input tokens, 55,000 cached tokens and 240 output tokens, model ID claude-haiku-5-5, endpoint https://opencode.ai/zen/go/v1/messages (@ai-sdk/anthropic) and 30-day data retention under Anthropic's data policy. Step 5 Preview Free has free input/output/cache read with unlimited requests and monthly allowance (limited-time; no end date given), model ID step-5-preview-free, endpoint https://opencode.ai/zen/go/v1/chat/completions. The supported-models, token-price, request-limit, endpoint and data-retention tables all gained rows for the two models, while no other model's prices, allowances or request limits changed

• 2026/10/08: Volcengine Ark Agent Plan (Personal Edition) widened the scope of its daily limit (reset at 00:00:00 every day) from "visual models only" to "visual models, speech models and Harness only" (this matches what the KB already says, so no substantive rule change); the announcement "Agent Plan package free ArkClaw benefit adjustment notice" was swapped for "Agent Plan Personal Edition & Professional Dataset joint promotion". Plan prices, tier allowances and all model deduction coefficients are unchanged

• 2026/10/06: OpenCode Go's Space Bunny Free limited-time free offer ended and the model became the regular paid model Space Bunny: token prices are input $0.15, output $0.60 and cache read \(0.03 per 1M tokens (identical on Go and Go Plus), with a monthly usage allowance of **\)30 on Go and $120 on Go Plus (no longer unlimited); estimated request limits are 3,130 / 5 hours, 7,810 / week and 15,630 / month on Go, and 12,500 / 5 hours, 31,250 / week and 62,500 / month on Go Plus, with a per-request usage estimate of 1,000 input tokens, 55,000 cached tokens and 200 output tokens. The model ID changed from space-bunny-free to space-bunny, the endpoint is unchanged (https://opencode.ai/zen/go/v1/chat/completions) and the data-retention table was renamed accordingly; the model's "limited-time" marker and the "Space Bunny Free:** limited-time free." note were removed, while no other model's prices, allowances or request limits changed

• 2026/10/03: Alibaba Cloud Bailian added the image model qwen-image-2.1-pro (text-to-image and image-to-image / image editing, billed per image), costing 0.25 CNY per image on the China site and the "Global" deployment (10 free images) and 0.283404 CNY per image internationally; the qwen-image-2.0 series (qwen-image-2.0-pro and its 2026-06-22/2026-04-22/2026-03-03 versions at 0.5 CNY per image, qwen-image-2.0 and its 2026-03-03 version at 0.2 CNY per image; internationally 0.550443 / 0.256873 CNY per image; 100 free images each) keeps its unit prices and free quotas unchanged. This change only concerns image-generation pricing in Bailian's general model catalogue and alters no coding plan's model library, tiers or quotas, the Token Plan included
• 2026/09/30: MiniMax launched M Plan to succeed Token Plan (announcement dated 2026-09-29): M Plan has three tiers — Go (¥49 / $22), Explore (¥119 / $55) and Build (¥469 / $132) — with the M3.1 Flash Preview text model and H3 video on Explore/Build; annual billing costs 10 months for 12; the first month is 50% off during the launch (through 2026-10-14); 1,000 credits = ¥7 (international $1). Token Plan is no longer sold; existing subscribers keep their plan and benefits if auto-renewal stays on (retaining historical perks such as unlimited quotas and the 150% weekly quota), and can upgrade to M Plan (Plus → Go/Explore/Build, Max → Explore/Build, Ultra → Build); after upgrading you cannot return to Token Plan
• 2026/09/30: Tencent Cloud LLM Token Plan expanded and reshuffled its model library: the Enterprise Professional Plan (Guangzhou/Singapore) added three models — Hy4 preview, MiMo-V2.6-Pro and MiMo-V2.6-Flash — removed the flat-credit-priced deepseek-v4-flash (the variant without the "0731 Official" suffix), and completed the two Official direct-supply model names to "DeepSeek-V4-Flash 0731 Official direct-supply" and "DeepSeek-V4-Pro 0813 Official direct-supply" (model IDs unchanged); credit prices were published for the three new models — Hy4 preview cached input 30 / uncached input 600 / output 1800 (identical in both regions, estimated blended price ~158, 500K credits covering roughly 3.165B tokens), MiMo-V2.6-Pro 2.5/300/600 in Guangzhou and 2.59/313.05/626.1 in Singapore (blended ~107 / ~112), MiMo-V2.6-Flash 2/100/200 in Guangzhou and 2.02/100.75/201.5 in Singapore (blended ~33), all in credits per 1M tokens with no peak/off-peak distinction. The Personal Edition Universal Token Plan added DeepSeek-V4.1-Flash Official direct-supply (model ID deepseek/deepseek-flash) and MiMo-V2.6-Flash to its supported models, with tier credit allowances and the Hy Token Plan unchanged. On peak-valley billing, the Singapore region is now unified to the same wording as Guangzhou ("peak hours are weekdays (Mon–Fri) 9:00–12:00 and 14:00–18:00 with all other times off-peak, weekends billed at off-peak rates all day; effective for Official direct-supply models from 2026-08-29 00:00 and for other DeepSeek models except deepseek-v4-flash and deepseek-v4-pro from 2026-09-26 00:00"); in the Guangzhou price table the Auto row was renamed from "Auto intelligent routing" to "Auto model" (matching Singapore), and the DeepSeek disclaimer changed from enumerating specific model names to "model services supplied directly by DeepSeek". Peak/off-peak unit prices, all other model credit prices and plan prices are unchanged
• 2026/09/30: MiniMax renamed "Token Plan" to "M Plan" in its official documentation: every "Token Plan" mention in the official docs (models overview and pay-as-you-go pricing, four pages across the Chinese and English sites) has been replaced by "M Plan" (e.g. "available only through M Plan and MiniMax Code for now", the "M Plan MCP" plugin, and "when API-vlm is called through M Plan, usage deducts from the included M Plan quota according to its pay-as-you-go price"), and the docs site's Pricing menu and documentation tab bar now list M Plan (pointing to /docs/m-plan/intro) with the old Token Plan tab hidden and the external link changed to https://www.minimax.cn/m-plan; the Token Plan FAQ and Token Plan intro pages still say "Token Plan" (as does the pay-as-you-go page's "resource coverage is the same as Token Plan" sentence and its link to the Token Plan pricing page), showing the rename was still in progress. (Re-checked on 2026-09-30: M Plan is in fact a new tier structure succeeding Token Plan — see the entry above.)
• 2026/09/29: Volcano Engine extended the kimi-k2.8-preview 40%-off deduction promotion window from 2026-09-30 23:59:59 to 2026-10-14 23:59:59: in the Agent Plan the limited-time 40% off (4.8) on the model's deduction coefficient of 8 now runs until the later date (start 2026-09-17 00:00 unchanged); in the Coding Plan the window during which this model's available quota is "comparable to that during the Agent Plan 40%-off deduction promotion" likewise moved from 2026-09-18 00:00 - 2026-09-30 23:59 to 2026-10-14 23:59. The Agent Plan (Personal Edition) billing, Agent Plan (Personal Edition) overview and Coding Plan (Personal Edition) overview were updated together; plan prices, tier quotas and every other model's deduction coefficient are unchanged
• 2026/09/29: The Alibaba Cloud Bailian Token Plan Personal Edition added one sentence below the upgrade proration formula — "the actual prorated amount is subject to the price breakdown shown on the payment page" — while the proration formula, the worked example and all tier prices are unchanged
• 2026/09/29: The Usage limits section of the OpenCode Go doc (zh-cn) dropped its illustrative example: the paragraph "e.g. if a model's monthly limit on Go is $60 and on Go Plus $120 …" and the sentence "when counting across models, the 5-hour, weekly and monthly allowances on Go are $12, $30 and $60 respectively; on Go Plus they are $48, $120 and $240" were removed entirely, leaving only "each model's 5 hours = 20% of the monthly limit, week = 50%, month = 100%" and "each model's monthly limit below determines how its usage counts toward these allowances"; the per-model monthly allowance table, the token price table and the request limit table are unchanged, so the cross-model totals are no longer stated in the doc
• 2026/09/28: The Volcengine Agent Plan Personal edition overview and billing doc (AFP deduction rules) mark doubao-seedream-5.0-lite as "being retired" (both the model-table row and the model name in the AFP deduction example; every tier is still ticked, and no retirement date has been given by the vendor)
• 2026/09/28: The MiniMax Token Plan FAQ gained two new sections ("Subscription Entitlement Adjustment" and "Token Plan Migration Guide"), the first official documentation of the tier migration and compensation rules: Max-hs (\(80/month) and Ultra-hs (\)150/month) are retired and migrate to Max / Ultra (monthly fee down $30 in both cases, with roughly $60 of equivalent credits granted each month; annual plans receive those grants month by month, each valid for 1 year and non-rolling); Starter and Plus-hs become existing-subscriber-only legacy tiers and are no longer sold to new users; benefits promised to users who subscribed before 2026-06-05 remain in effect and the validity of compensation credits is corrected automatically from 1 month to 1 year; Plus/Max keep their prices while M2.7 5-hour usage rises about 10% and M3 access plus multimodal quota are added; credits can be used across most platform models (MiniMax H3 not supported yet) and are shared across modalities. Note that the FAQ's migration section still quotes Plus $20 / Max $50 / new Ultra $120, which does not match the \(22/\)55/$132 on the international pricing page (the FAQ was not updated for the 2026-08-26 price increase)
• 2026/09/28: OpenCode Go added a **Go Plus (\(40/month)** plan alongside Go (\)10/month): token prices are identical on both plans and Go Plus only raises the per-model usage limits — each model's monthly allowance is generally 2–8x that of Go (mostly 4x, e.g. GLM-5.3 \(15→\)120, GLM-5.3-Flash \(60→\)180, Kimi K2.6 \(60→\)240, DeepSeek V4 Flash \(30→\)120), while the cross-model totals are \(12/\)30/$60 (5 hours/week/month) on Go and \(48/\)120/$240 on Go Plus; the subscription entry point was renamed from OpenCode Zen to OpenCode Console, and each workspace still allows only one member to subscribe to Go or Go Plus. The same update removed GLM-5.1, Qwen3.7 Max, Qwen3.6 Plus and MiniMax M2.5 from the supported models (Omen Alpha is no longer supported either; the kb no longer lists it as a supported model)
• 2026/09/27: OpenCode Go added the limited-time free LongCat 2.5 Preview Free model: input, output and cache read are all Free, with the request limits and the monthly usage allowance both listed as unlimited (a limited-time promotion with no end date given by the vendor); model ID longcat-2.5-preview-free, endpoint https://opencode.ai/zen/go/v1/chat/completions (the supported model list, token price table, request limit table, endpoint table and data retention table all gained this model; pricing and allowances of other models are unchanged)
• 2026/09/26: The 4x boost of the OpenCode Go DeepSeek V4.1 Flash quota has changed from a limited-time promotion into a standard value: in the pricing table the model's monthly usage allowance changed from "~~\(15~~ **\)60 4x · ends Sep 27" to a plain $60, and the estimated-request table likewise changed "~~6,500~~ 26,000 (4x · ends Sep 27)" to 26,000 / 65,000 / 130,000** — the old $15 and 6,500/16,250/32,500 baselines and the promotion end-date annotation have been removed entirely, i.e. the boosted allowance and the raised request limits are now the model's regular quota; token pricing unchanged (off-peak input $0.15/1M, output $0.60/1M, cache read $0.003/1M; peak pricing is double)
• 2026/09/25: The Infini (Wuwen Xinqiong) GenStudio changelog published a "2026-10-09 release preview": deepseek-v4-flash will be retired at 12:00 on 9 October 2026, with deepseek-v4.1-flash recommended as the replacement model — migrate in advance (this platform's Infini Coding Plan was already discontinued on 2026-06-27; this change affects GenStudio API calls)
• 2026/09/25: Alibaba Cloud Bailian added the domain model decision-model-preview and brought it into the Token Plan Personal Edition: model pricing gained a "decision models" section (billed on input tokens, with decision-model-preview marked "limited-time free" as its input price, covering both China North 2 (Beijing) and Singapore), the "Choose a model" list gained a "decision models" category described as "a structured decision model for high-frequency business judgement that completes classification, yes/no decisions and scoring in a single forward pass and returns a probability distribution and confidence" (model detail page /zh/model-studio/decision-model-preview), and the Token Plan Personal Edition added to its benefits "decision-model-preview is free for a limited time: calls consume no Credits, see Connecting to the decision model for setup" plus a new row in the supported-model table "domain model / decision-model-preview / decision model"
• 2026/09/25: The Tencent Cloud Token Plan Enterprise Professional Plan rewrote its DeepSeek peak-valley billing rule and added a non-Official-direct-supply DeepSeek-V4.1-Flash: the notes changed from "Official direct-supply models are billed at off-peak rates all weekend; Official models have peak hours Mon–Sun" to a unified "all DeepSeek models have peak hours on weekdays (Mon–Fri) 9:00–12:00 and 14:00–18:00 with all other times off-peak, and weekends (Sat-Sun) are billed at off-peak rates all day", plus new effective dates — Official direct-supply models from 2026-08-29 00:00 Beijing time, and other DeepSeek models (except deepseek-v4-flash and deepseek-v4-pro, which are billed at flat credits and have no peak/off-peak split) from 2026-09-26 00:00 Beijing time; that means platform-hosted models such as DeepSeek-V4-Flash 0731 Official and DeepSeek-V4-Pro 0813 Official also get the all-day off-peak weekend rate from 26 September. This update was applied only to the Guangzhou region; the Singapore region still carries the old wording. The model library (both Guangzhou and Singapore) also added the non-Official-direct-supply DeepSeek-V4.1-Flash (model ID deepseek-v4.1-flash, coexisting with the official direct-supply deepseek/deepseek-flash), whose credit prices are identical to the Official direct-supply version — cached input 2 / uncached input 100 / output 400 (off-peak), 4 / 200 / 800 (peak) credits per 1M tokens, an estimated blended price of ~26 / ~51 credits per 1M tokens
• 2026/09/24: Alibaba Cloud Bailian added a peak/off-peak time definition to its DeepSeek third-party models: "busy hours are 08:00–22:00 Beijing time, idle hours are 22:00–08:00 the next day" — i.e. the DeepSeek models whose prices are split into peak/off-peak are charged with 08:00–22:00 as peak and 22:00–08:00 as off-peak each day (the same window as the Bailian Token Plan Personal Edition night discount). No prices or other billing rules changed
• 2026/09/24: OpenCode Go added GPT 6 Luna to its model list: billed in 272K-token tiers — up to 272K tokens input $0.10 / output $0.50 / cache read $0.01 / cache write $0.125 per 1M tokens, above 272K tokens \(0.20/\)0.75/\(0.02/\)0.25; the input and cache prices are exactly half of GPT 5.6 Luna's and the output price is lower (was $1.20), while GPT 5.6 Luna stays in both the model list and the pricing table. Monthly usage allowance $15, request limits 4,230 per 5 hours, 10,560 per week, 21,130 per month; model ID gpt-6-luna, endpoint https://opencode.ai/zen/go/v1/responses (the data-retention note was updated from "GPT 5.6 Luna" to "GPT 6 Luna / GPT 5.6 Luna")
• 2026/09/23: The Zhipu GLM Coding Plan "GLM-5.3-Flash Usage Campaign" extended its eligible clients from ZCode to ZCode and AutoClaw (autoclaw.zhipuai.cn): plan users get unlimited GLM-5.3-Flash usage on both clients, with the campaign period and other terms unchanged (2026-09-03 to 2026-10-07, daily 23:00–09:00; doubled quota on other agents)
• 2026/09/23: OpenCode Go added the limited-time free Space Bunny Free model: input, output and cache read are all Free, with the request limits and monthly usage allowance both listed as unlimited; model ID space-bunny-free, endpoint https://opencode.ai/zen/go/v1/chat/completions (the vendor gave no end date for this limited-time promotion)
• 2026/09/23: The Volcano Engine Coding Plan (Personal Edition) and Agent Plan (Personal Edition) model libraries were updated: added doubao-seed-2.1-pro (new-generation flagship model with all-round capability gains, 1024k context window / 256k output), doubao-seed-2.1-lite (lightweight and efficient, 1024k context window / 256k output), doubao-seed-2.0-mini (ultra-fast responses, 256k context window / 128k output) and deepseek-v4.1-flash (the lightweight flagship of DeepSeek's new-architecture series, 552B total-parameter MoE, natively capable of multimodal vision understanding, 1024k context window / 384k max output; it had already joined the Agent Plan and now joins the Coding Plan); doubao-seed-2.1-turbo and doubao-seed-2.0-lite are marked "being retired" (the official announcement calls it "model retirement initiated"), and doubao-seedance-1.5-pro was removed from the Agent Plan model table; the Agent Plan deduction coefficients were updated accordingly — doubao-seed-2.1-pro 2.5, doubao-seed-2.1-lite 0.5 (4.5 with audio input), and doubao-seed-2.0-mini refined to 0.25 without audio input / 2.5 with audio input; both the Coding Plan and Agent Plan 1M-context supported lists now include deepseek-v4.1-flash (Agent Plan list: doubao-seed-evolving, glm-5.3, glm-5.3-flash, kimi-k3, kimi-k2.8-preview, deepseek-v4.1-flash, deepseek-v4-flash, deepseek-v4-pro)
• 2026/09/23: The Volcano Engine deepseek-v4.1-flash limited-time 50%-off deduction window was extended and expanded to the Coding Plan: in the Agent Plan the promotion deadline moved from 2026-09-28 23:59 to 2026-10-30 18:00 (start unchanged at 2026-09-15 00:00, deduction coefficient 2.5 → 1.25), and the Coding Plan added the same promotion for 2026-09-23 00:00 to 2026-10-30 18:00
• 2026/09/23: The CTCloud Programming Token Plan switched to credit metering and gained a "Programming Token Plan (Credits Edition)": tiers and prices are unchanged (29/89/199/399/699 RMB per month) but the quota changed from a fixed token allowance to 3,000/10,000/25,000/50,000/100,000 credits per subscription month, and the supported models became DeepSeek-V4-Pro, DeepSeek-V4-Flash-0731, GLM-5.2, GLM-5.1, Kimi-K2.6 and MiniMax-M3 (GLM-5.0 and DeepSeek-V3.2 moved out). A "credit-to-model conversion" table was added (tokens per credit, input/output — DeepSeek-V4-Pro 1,111/370, DeepSeek-V4-Flash 3,333/1,111, GLM-5.2 1,250/357, GLM-5.1 input [0,32k] 1,667 (output 417) and input (32k,200k] 1,250 (output 357), Kimi-K2.6 1,538/370, MiniMax-M3 input [0,512k] 4,762 (output 1,190) and input (512k,1M] 2,381 (output 595)) plus a note that the model library is updated dynamically and no given model is guaranteed permanently. The previous token-metered plans are labelled "old plans being retired — only existing subscribers may renew, new purchases not supported", with their supported models now DeepSeek-V4-Flash-0731, GLM-5.1, GLM-5.0 (retiring 2026-10-10) and DeepSeek-V3.2 (written "DeepSeeV3.2" by the vendor, retiring 2026-10-10)
• 2026/09/23: The Kuaishou Wanqing (StreamLake/Vanchin, the platform behind the KwaiKAT Coding Plan) model list added DeepSeek-V4.1-Flash (listed in the multimodal section: deep thinking + image understanding, 1M context / 384K max output, native multimodal visual understanding, default rate limit RPM 10 / TPM 300,000); Wanqing's pay-as-you-go model catalog does not state whether the model is covered by the KwaiKAT Coding Plan quota
• 2026/09/23: Alibaba Cloud Bailian Token Plan (Personal Edition) abolished the weekly limit and switched to a subscription-month quota: the Personal Edition moved from a "fixed 7-day window quota" to a monthly quota — a subscription month runs 30 days from the subscription date, and once cumulative consumption within the subscription month reaches the plan quota, service is paused until the quota resets in the next subscription month (auto-refreshed based on the subscription date, not a fixed calendar date), with unused quota not rolling over; per-tier quotas were adjusted to Lite 11,500 / Essential 25,500 / Standard 45,000 / Pro 180,000 Credits per month (tiers and prices unchanged, limited-time prices 39/79/139/499 RMB per month), and the remaining quota of existing subscriptions was reset once to the full monthly quota of the corresponding plan on 2026-09-22 with the subscription cycle and expiry date unchanged; the previous "reset card / quota reset entitlement" (the "Reset quota" button and reset-count notes) was removed entirely, with quota resets now handled automatically each subscription month; the deduction order became "each request is deducted first from the current subscription month's plan quota, and once exhausted, from Usage Pack quota", and Usage Packs are no longer described as "not subject to the 7-day window limit" but rather as "neither occupying nor counting toward the plan monthly quota"; the upgrade rule changed so that the current subscription cycle is not reset — the price difference is paid as (new plan price − old plan price) × remaining days ÷ 30 (partial days counted as a full day), and the new quota for the current cycle is granted as remaining days ÷ 30 × (new monthly quota − old monthly quota), rounded up, with the new tier's monthly quota applying from the next subscription month; the 429 error message changed from insufficient_quota: Your token-plan 1-week quota has been exhausted. to insufficient_quota: Your token-plan quota has been exhausted.; the Team Edition FAQ comparison table's "Personal Edition quota mechanism" was updated to "monthly quota, auto-refreshed per subscription month", and the Team Edition overview table dropped the "7-day limit: unlimited" row
• 2026/09/22: The SenseTime SenseNova API docs (including the TokenPlan credits rules and model overview) changed their model library: DeepSeek V4 Pro was removed from the model overview and DeepSeek V4.1 Flash was added (model ID deepseek-flash, 1M context, image input support), while DeepSeek V4 Flash (deepseek-v4-flash) stays listed. The corresponding model sections were rewritten and gained an image-input section with limits (single image ≤50 MB, ≤200 images per request, ≤64 MB per request body, ≤200 MB total for URL-passed images; video input is only supported by the Chat Completions API), max_tokens now defaults to 131072 (range 1–393216, thinking is truncated when it exceeds the limit), reasoning_effort has native none/low/high/max levels with compatibility mappings, and explicit caching is stated as unsupported
• 2026/09/22: Zhipu GLM Coding Plan (announced on both the China site bigmodel.cn and the international z.ai DevPack) added two limited-time campaigns: the "GLM-5.3-Flash Usage Campaign" — from 2026-09-03 to 2026-10-07, daily 23:00–09:00, plan users get unlimited GLM-5.3-Flash usage via ZCode plus doubled quota on other agents; and the "Double Festival" campaign — from 2026-09-25 to 2026-10-07 all-day usage is charged at the off-peak rate (50% points consumption around the clock). The vendor notes that stacking the night campaign on top makes the actual available quota far higher than the allowance reference table suggests
• 2026/09/22: Alibaba Cloud Bailian model calling prices, model library and context cache updated: the Stepfun deployment added stepfun/step-5-preview (7 CNY input / 20 CNY output per million tokens, no free quota, alongside stepfun/step-3.7-flash at 1.35/8.1 CNY) and the Unisound section lists unisound/unisound-u2 (1 CNY input / 2 CNY output); a new "audio generation" billing section was added (qwen-audio-3.1-tts-next, billed on input and output tokens at 6/12 CNY); speech recognition added the Qwen-Audio-3.1 series and switched from per-second audio billing to per-token billing — qwen-audio-3.1-asr-flash-message and qwen-audio-3.1-asr-flash-streaming cost 6/4.5 CNY (international 6.781/5.104 CNY), qwen-audio-3.1-asr-flash-filetrans and qwen-audio-3.1-asr-flash cost 0.8/2.7 CNY (international 1.094/3.427 CNY), while the 3.0 series per-second tiers remain; the "Choose a model" list switched its speech-recognition recommendation from qwen-audio-3.0-asr-flash-streaming/-filetrans to the 3.1 versions; the context-cache Stepfun (StepFun deployment) list added stepfun/step-5-preview; qwen3.8-max-prime's fast-mode doc link changed from /fast-mode to /prime-mode; the Workspace ID link in the official OpenCode / OpenClaw integration docs changed from "Get Workspace ID" to the regions page (/zh/model-studio/regions#h2_migrate_domain)
• 2026/09/22: Alibaba Cloud Bailian Token Plan (Personal Edition) night discount changed: qwen3.8-max went from 50% off to 40% off and qwen3.8-flash was added at 40% off; deepseek-v4-pro-0813, deepseek-v4-flash-0731 and deepseek-v4.1-flash stay at 50% off (all are Credits-consumption discounts for calls between 22:00 and 08:00)
• 2026/09/22: OpenCode Go added three models — Grok 4.7, MiMo-V2.6-Flash and MiMo-V2.6-Pro. Grok 4.7 has the same pricing and request limits as Grok 4.6 (≤200K tokens: input $2.00 / output $6.00 / cache read $0.50; >200K tokens: \(4.00/\)12.00/$1.00 per 1M tokens; allowance \(15/month; request limits 169 per 5 hours, 423 per week, 845 per month; model ID grok-4.7). MiMo-V2.6-Flash/V2.6-Pro have pricing and request limits identical to the previous generations they replace, MiMo-V2.5/V2.5-Pro (\)0.14/\(0.28/\)0.0028 with a $60 allowance, and \(0.435/\)0.87/$0.003625 with a $15 allowance)
• 2026/09/22: The Xiaomi MiMo Token Plan was reworked into the "Personal Edition" and its supported models went from 9 to 8 — mimo-v2.6-pro and mimo-v2.6-flash were added (Credit conversion matches the previous generation: V2.6-Pro 2.5 cached / 300 uncached / 600 output, V2.6-Flash 2/100/200 Credits per token), while MiMo-V2-Pro, MiMo-V2-Omni and MiMo-V2-TTS were removed from the supported-model list and the Credit table; a notice states that mimo-v2.5-pro and mimo-v2.5 will be retired at 10:00 Beijing time on 2026-10-21. Plan prices and quotas are unchanged (39/99/329/659 RMB or 6/16/50/100 USD per month, 4.1B/11B/38B/82B Credits per month). The discount description was narrowed to "88% off first purchase, 88% off annual auto-renewal, 0.8x consumption at night", the earlier Token Plan upgrade "Credits usage reset" campaign (effective 2026-05-27) was removed, and the purchase note now reads "only one Personal Edition plan can be held at a time"
• 2026/09/21: Alibaba Cloud Bailian Token Plan (Personal Edition) added auto to its supported models (official description: "a platform-provided smart model that automatically matches the underlying model to the request, balancing quality and cost"; capability label "reasoning, text generation"; listed first under the Qwen brand in the model table). The official OpenCode / OpenClaw integration docs added an auto entry to both the Personal and Team Edition configurations (in OpenCode it only declares text input/output, with no reasoning or limit fields; in OpenClaw contextWindow 1,000,000, maxTokens 393,216, reasoning false) and changed OpenClaw's default model from bailian-token-plan/qwen3.8-flash to bailian-token-plan/auto
• 2026/09/21: Alibaba Cloud Bailian's model calling prices and context-cache supported model list added two models — ZHIPU/GLM-5.3-FlashX (Zhipu deployment, thinking mode only, 2 CNY input / 7 CNY output per million tokens, no free quota; cache-hit discount 28.5%, higher than the 25% of the other five models in the same series such as ZHIPU/GLM-5.3-Flash) and vanchin/deepseek-v4.1-flash (Kuaishou Wanqing deployment, 2 CNY input / 8 CNY output per million tokens; cache-hit discount 2%, lower than vanchin/deepseek-v4-pro's 8.33%)
• 2026/09/21: OpenCode Go extended its limited-time 4x quota boost for DeepSeek V4.1 Flash — the end date moved from 2026-09-20 to 2026-09-27 (the $60 monthly usage allowance, the request limits of 26,000 per 5 hours / 65,000 per week / 130,000 per month and the token pricing are all unchanged)
• 2026/09/20: StepFun Step Plan added step-5-preview to its supported models (next-generation flagship base model for real-world tasks) and removed step-image-edit-2 along with the previous image model retirement notice, so the plan no longer covers any text-to-image or image editing model; the official model description correspondingly changed from "flagship models covering text, reasoning, voice, image editing and intelligent routing" to "text, reasoning, voice and intelligent routing". The official docs also split the Base URL guidance per tool — Claude Code / Anthropic SDK use https://api.stepfun.com/step_plan while OpenAI SDK Chat Completions calls still use https://api.stepfun.com/step_plan/v1 — and clarified that the Step Plan channel consumes plan Credits and is independent of the regular API channel quota (the FAQ now likewise says to pick the address by tool)
• 2026/09/18: The Kimi Code docs added per-model ID membership tier requirements and introduced the Plus / Pro tier names for the first time — k3 / k3-256k require Moderato / Plus or above (1M context requires Allegretto / Pro or above); the wording for kimi-for-coding changed from "available to all members" to "available to Andante / Plus or above"; kimi-for-coding-highspeed requires Allegretto / Pro or above. It also added overseas API endpoint domains: Kimi Code overseas OpenAI-compatible https://api.kimi.ai/coding/v1 and Anthropic-compatible https://api.kimi.ai/coding/, and Kimi Open Platform overseas https://api.moonshot.ai/v1 (domestic endpoints are api.kimi.com/coding and api.moonshot.cn/v1)
• 2026/09/18: OpenCode Go removed the Union Alpha Free model (limited-time free period ended) — the model (model ID union-alpha) was removed from the supported model list, the token price table, the request limit table and the endpoint table
• 2026/09/18: Volcano Engine Coding Plan (Personal Edition) and Agent Plan (Personal Edition) added the kimi-k2.8-preview model (overall performance close to K3 with higher thinking efficiency, supports text and image input; 1M context window / 1M max output, supported by all plans); Agent Plan deduction coefficient 8, at 40% off (4.8) from 2026-09-17 00:00 to 2026-09-30 23:59; during the promotion from 2026-09-18 00:00 to 2026-09-30 23:59 the available quota of this model in the Coding Plan is comparable to that during the Agent Plan 40%-off deduction promotion; the model also joins the 1M-context supported list (now glm-5.3, glm-5.3-flash, deepseek-v4.1-flash, deepseek-v4-flash, deepseek-v4-pro, kimi-k3, kimi-k2.8-preview)
• 2026/09/18: Alibaba Cloud Bailian Token Plan Personal Edition added the Essential tier, expanding the personal tiers from three (Lite/Standard/Pro) to four: list price 120 RMB/month, 79 RMB/month for a limited time, 5,625 Credits every 7 days (2.25x Lite), supports 2-3 concurrent Agents, and its benefits are all Lite benefits; the vendor also moved the Usage Pack (100 RMB each/month, 20,000 Credits; requires an active subscription, up to 5 held at a time, not subject to the 7-day window limit) from a table row to a footnote
• 2026/09/17: OpenCode Go added the Union Alpha Free model (limited time): input/output/cache read are all Free, with unlimited requests and quota (limited time); model ID union-alpha, endpoint https://opencode.ai/zen/go/v1/messages (Anthropic-compatible)
• 2026/09/16: Alibaba Cloud Bailian Token Plan (Personal Edition) added glm-5.3 to its "Supported models" list (Zhipu AI, capability label "reasoning, text generation", coexisting with glm-5.2; glm-5.3 had already been added to the Bailian model catalog on 2026-09-15 but was not part of the Token Plan at that time); the official integration docs added the corresponding model config — OpenClaw with contextWindow 1,000,000, maxTokens 16,384, reasoning false, and OpenCode with thinking enabled (budgetTokens 8192); the Personal Edition limited-time night 50%-off list is unchanged (qwen3.8-max, deepseek-v4-pro-0813, deepseek-v4-flash-0731, deepseek-v4.1-flash)
• 2026/09/15: Kuaishou Wanqing (StreamLake/Vanchin, the platform behind the KwaiKAT Coding Plan) model list added three models — GLM-5.3, GLM-5.3-Flash and DeepSeek-V4-Pro-0813; Wanqing's pay-as-you-go model catalog does not state whether these models are covered by the KwaiKAT Coding Plan quota
• 2026/09/15: Alibaba Cloud Bailian model calling prices added the GLM model glm-5.3 (non-thinking and thinking modes, no token tiers, 8 CNY input / 28 CNY output per million tokens, the same price for China-site regions such as China North 2 (Beijing) and the "global" deployment; 10.208/32.084 CNY on the international site) with 1 million free tokens; the list of GLM models supported by context caching (deployed on Alibaba Cloud Bailian) also added glm-5.3 (25% cache-hit discount, the same tier as glm-5.2 and glm-5.2-fast-preview); meanwhile the mode of ZHIPU/GLM-5.3 (GLM deployed by Zhipu) was corrected from "non-thinking and thinking modes" to "thinking mode only" (price unchanged at 8 CNY input / 28 CNY output, no free quota)
• 2026/09/15: Volcano Engine Agent Plan (Personal Edition) added the deepseek-v4.1-flash model (supported by all plans; 1M context window / 384K max output, natively capable of multimodal vision understanding) with a deduction coefficient of 2.5, at 50% off (1.25) from 2026-09-15 00:00 to 2026-09-28 23:59; the model also joins the 1M-context supported list (now glm-5.3, glm-5.3-flash, deepseek-v4.1-flash, deepseek-v4-flash, deepseek-v4-pro, kimi-k3)
• 2026/09/14: In the Alibaba Cloud Bailian Token Plan (Personal Edition) supported-model list, deepseek-v4.1-flash's capability label changed from "reasoning, text generation" to "reasoning, vision understanding, text generation"; the official integration doc (OpenClaw) also changed that model's config input from ["text"] to ["text", "image"], confirming the model accepts image input directly within the Token Plan
• 2026/09/14: Volcano Engine Agent Plan (Personal Edition) added two multimodal models: doubao-seedream-5-0-pro (image generation, all plans; input images are free for the first image and 10 AFP per image thereafter, output images cost 150 AFP for ≤2.61M pixels and 300 AFP above that for single-image generation, and 75/150 AFP for layer separation) and doubao-seedance-2.5 (video generation, Large/Max; 210 per token when the input contains video and 350 when it does not for 480p/720p, 230/385 for 1080p); doubao-seedance-1.5-pro is marked as being retired
• 2026/09/14: Alibaba Cloud Bailian Token Plan (Personal Edition) added deepseek-v4.1-flash to its supported models (it had previously only been added to the Bailian model library, not to the Token Plan); the model also joined the Personal Edition limited-time night discount (50% off Credits from 22:00 to 08:00), whose list is now qwen3.8-max, deepseek-v4-pro-0813, deepseek-v4-flash-0731, deepseek-v4.1-flash
• 2026/09/14: Tencent Cloud Token Plan Enterprise Professional Plan (Guangzhou/Singapore) added DeepSeek-V4.1-Flash Official direct-supply (model ID deepseek/deepseek-flash) to its model library; the Official direct-supply Flash-series credit prices were lowered across the board — DeepSeek-V4-Flash 0731 Official direct-supply (previously ~39 off-peak / ~77 peak credits per 1M tokens) and DeepSeek-V4-Flash-Vision-Exp Official direct-supply (previously ~35 / ~70) both dropped to match the new model (cached 2 / uncached 100 / output 400 off-peak, 4/200/800 peak; estimated blended ~26 / ~51), following DeepSeek's upstream move of routing the legacy model names to V4.1-Flash billing; it also removed the GLM-5.3-Flash 50%-off credit-price promotion that had expired on 2026-09-10
• 2026/09/14: OpenCode Go temporarily boosted the DeepSeek V4.1 Flash limits 4x: monthly usage allowance raised from $15 to $60 (promotion ends 2026-09-20), request limits raised from 6,500 per 5 hours / 16,250 per week / 32,500 per month to 26,000 per 5 hours / 65,000 per week / 130,000 per month; token pricing unchanged
• 2026/09/14: Alibaba Cloud Bailian added deepseek-v4.1-flash to its model library: China North 2 (Beijing) / global pricing is peak input 2 RMB, off-peak input 1 RMB, peak output 8 RMB, off-peak output 4 RMB per 1M tokens (with a 1M-token free quota); international site peak 2.188/8.75 RMB, off-peak 1.094/4.375 RMB. Context-cache hits cost 10% of the input price (peak 0.2 RMB, off-peak 0.1 RMB), a better discount than the 20% applied to other Bailian models but still pricier than DeepSeek's official API cache-hit price (0.04/0.02 RMB). The "Choose a model" list also swapped its featured slot from deepseek-v4-flash-0731 to this model; it is not yet in the Bailian Token Plan (Personal/Team Edition) supported-model lists
• 2026/09/12: OpenCode Go renamed the DeepSeek V4.1 Flash model ID from deepseek-flash to deepseek-v4.1-flash
• 2026/09/12: Kimi Code's standard tier moved from K2.7 Code to K2.8 Preview (model name kimi-for-coding now maps to K2.8 Preview, supporting up to 1M context and low/high/max thinking levels); the High Speed tier remains K2.7 Code HighSpeed
• 2026/09/12: DeepSeek reversed its V4 Pro retirement plan: the vendor will keep providing the DeepSeek V4 Pro API service after 2026-09-14 with unchanged billing
• 2026/09/10: Added DeepSeek-V4.1-Flash to the model parameters comparison table (552B total parameters, 8B input / 16B output activation, multimodal support)
• 2026/09/10: DeepSeek released DeepSeek-V4.1-Flash (new model name deepseek-flash, 1M context, supports image understanding) with significantly lower pricing (off-peak: cached 0.02, uncached 1, output 4 RMB; peak 0.04/2/8 RMB per 1M tokens); old model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are discontinued (requests served by V4.1-Flash at Flash pricing); V4 Pro is planned for orderly retirement — after 2026-09-14 12:00 all deepseek-v4-pro requests route to V4.1 Flash at Flash pricing
• 2026/09/10: Zhipu GLM-5.3-Flash limited-time 50% off promotion ended, reverting to standard pricing (uncached input 0.8, output 2.8, cached input 0.23 RMB per 1M tokens); Alibaba Cloud Bailian also removed the "50% off (limited time)" label for the same model; Tencent Cloud Token Plan (Personal Edition Universal Plan and Enterprise Professional Plan) marked GLM-5, GLM-5.1 and GLM-5-Turbo for discontinuation on 2026-10-09; OpenCode Go billing limits redefined as per-model monthly amounts (5 hours = 20% of monthly, week = 50%, month = 100%), with differing monthly limits per model (e.g. GLM-5.3 $15, GLM-5.3-Flash $60)
• 2026/09/09: Volcano Engine Coding Plan (Personal Edition) added the Kimi-K3 model (1M context / 128K max output, native visual understanding, high deduction coefficient, recommended only for Pro plan users); iFlytek Astron Token Plan Team Edition added the Spark-X2.5 model (256K, input 320 / cache 48 / output 1200 / thinking 1200 credits per 1M tokens), Spark-X2 and Spark-X2-Agent discontinued
• 2026/09/04: StepFun Step Plan announced that the step-image-edit-2 model will be retired on 2026-10-10, and the Step Plan text-to-image and image editing APIs will stop serving requests simultaneously
• 2026/09/04: OpenCode Go added support for the Omen Alpha model (input $0.20/1M, output $0.66/1M, cache read $0.04/1M, usage allowance $100/month; request limits 11,600 per 5 hours, 29,000 per week, 57,900 per month; model ID omen-alpha)
• 2026/09/04: Tencent Cloud LLM Token Plan expanded its model library: both the Universal Token Plan (Personal Edition) and the Enterprise Professional Plan added the GLM-5.3-Flash and Kimi K3 models; the Enterprise Professional Plan also published the credit prices for these two models (Guangzhou GLM-5.3-Flash 23/80/280, Kimi K3 200/2000/10000; Singapore GLM-5.3-Flash 21.5898/107.949/359.83, credits per 1M tokens)
• 2026/09/03: Kimi Code revised its "Core Advantages" section: it now explicitly states the plan is powered by the flagship K3 model (~2.8 trillion parameters) with the K2.7 Code in two modes (standard / High Speed), with a maximum inference speed of 260 Tokens/s (previously described as 100 Tokens/s) and a 1M Token context window; it removed the statements "High Speed version is 5–6x the standard version" and "approx. 300–1200 requests per 5 hours, max concurrency 30"
• 2026/09/03: Alibaba Cloud Bailian Token Plan (Personal Edition) usage policy tightened: the official guide changed from "you may configure the same API Key on multiple of your own devices (e.g., home and office computers)" to "for the subscriber's own use on a single device."
• 2026/09/03: OpenCode Go added support for the Muse Spark 1.3 Contributor model (Meta Contributor tier: allows Meta to use prompts and completions for training future models in exchange for heavily discounted token pricing; input $0.10/1M, output $0.20/1M, cache read $0.002/1M, usage allowance $60; request limits 45,300 per 5 hours, 113,300 per week, 226,600 per month; model ID muse-spark-1.3-contributor; only available in regions permitted by Meta's Geographic Use Policy)
• 2026/09/02: Tencent Cloud LLM Token Plan Personal Edition Hy Token Plan Pro Plan corrected its credits per subscription month from 1,560 to 4,760 (238 RMB/month price unchanged; 1,560 was a vendor mislabel/typo inconsistent with the price × 20 pattern of the other tiers)
• 2026/09/01: Volcano Engine Agent Plan (Personal Edition) deduction coefficient rule adjustment: the input and output deduction coefficients for text-generation/embedding models are now determined solely by the model and no longer vary with input length (previously input coefficient = model coefficient × input segment factor: ×0.67 for ≤32k, ×1 for 32k–128k, ×2 for >128k; the length segmentation has been removed). Input/output deduction coefficients per model: doubao-seed-2.0-mini 0.25, doubao-seed-2.0-lite/deepseek-v4-flash 0.5, glm-5.3-flash 0.5 (0.25 for the first two weeks at 50% off), doubao-seed-2.1-turbo/doubao-seed-evolving/minimax-m3 2.5, kimi-k2.7-code 4.5, glm-5.3 (glm-latest) 4.5, deepseek-v4-pro 5.5, kimi-k3 10, doubao-embedding-vision 0.5
• 2026/09/01: OpenCode Go Qwen3.7 Max model allowance reduced: request limits lowered from 340 per 5 hours, 840 per week, 1,690 per month to 170 per 5 hours, 420 per week, 840 per month, and the monthly usage allowance lowered from $60 to $30 (pricing input $2.50/1M, output $7.50/1M, cache read $0.50/1M, cache write $3.125/1M unchanged)
• 2026/08/31: Tencent Cloud LLM Token Plan switched to credit-based deduction: the Personal Edition changed from token quotas to credits effective 2026-08-31 17:00 (Universal Token Plan Lite/Standard/Pro/Max = 39/99/299/599 RMB/month, 780/1,980/5,980/11,980 credits/month; Hy Token Plan Lite/Standard/Pro/Max = 28/78/238/468 RMB/month, 560/1,560/1,560 (as-is in source doc, likely a typo)/9,360 credits/month); Universal Token Plan supported models updated to Auto, DeepSeek-V4-Flash/Pro Official direct-supply, MiniMax-M2.7, MiniMax-M3, GLM-5/5.1/5.2/5.3, Kimi K2.7 Code, Hy4 preview (Kimi-K2.5 discontinued), Hy Token Plan supports Hy3 and Hy4 preview (Hy3 preview auto-routes to Hy3); Enterprise Professional Plan added DeepSeek-V4-Flash-Vision-Exp Official direct-supply (multimodal/visual understanding, text capability on par with V4-Flash Official, multimodal Agent performance approaching Claude Opus-4.8) and removed Kimi-K2.5; DeepSeek V4 Official direct-supply peak-valley billing adjusted from 2026-08-29 to weekdays only (weekends all off-peak), DeepSeek V4 Official peak hours are Mon–Sun 9:00–12:00 and 14:00–18:00
• 2026/08/31: Volcano Engine Coding Plan (Personal Edition) and Agent Plan (Personal Edition) removed the GLM-5.2 model (previously marked as "phasing out"); GLM-5.2 is no longer listed among supported models or in the 1M-context supported list; the GLM-5.3 deduction coefficient description changed from "same as GLM-5.2" to "high"
• 2026/08/31: Alibaba Cloud Bailian model pricing added ZHIPU/GLM-5.3-Flash (thinking-only mode, input 0.8 RMB, output 2.8 RMB per 1M tokens, supports context-cache hit discount at 25%) and qwen-flash-character (0.25/1.5 RMB); qwen3-vl-rerank price lowered (text 0.7→0.5 RMB, image 1.8→0.5 RMB); Token Plan (Personal Edition) limited-time night discount added deepseek-v4-flash-0731
• 2026/08/29: OpenCode Go added the Hy4 preview model (input $0.834/1M, output $2.501/1M, cache read $0.042/1M, usage allowance $30; request limits 1,350 per 5 hours, 3,380 per week, 6,770 per month; model ID hy4-preview)
• 2026/08/29: Volcano Engine Coding Plan (Personal Edition) and Agent Plan (Personal Edition) added the glm-5.3-flash model (Zhipu's first natively multimodal model, 320B total params / 18B active, supports image input, 1M context / 128K max output; first two weeks deduction coefficient at 50% off, promotion ends 2026-09-11 23:59:59); Agent Plan's Agent Evolution changed to first 50 files free
• 2026/08/28: Zhipu International (Z.ai DevPack) GLM Coding Plan model change: all plans now support {GLM-5.3, GLM-5.3-Flash} (was {GLM-5.3, GLM-5-Flash}); requests for GLM-5.2/GLM-5.1 auto-routed to GLM-5.3, GLM-4.7 auto-routed to GLM-5.3-Flash (GLM-5-Turbo no longer a routing target); release notes removed the GLM-5V-Turbo and GLM-5-Turbo model entries
• 2026/08/28: OpenCode Go added support for Qwen3.8 Flash model (input $0.15/1M, output $0.47/1M, cache read $0.016/1M, cache write $0.20/1M, usage limit $30; request limits 5,400 per 5 hours, 13,500 per week, 27,000 per month; model ID qwen3.8-flash)
• 2026/08/28: Infini-AI GenStudio changelog: a batch of models were discontinued on 2026-08-28 (deepseek-r1, deepseek-v3 series, deepseek-v3.2/v3.2-thinking, glm-4.5/4.5-air/4.6/4.7/5/4.5v/4.6v, minimax-m2.1/m2.5, kimi-k2.5), recommended migration to deepseek-v4-flash/v4-pro, glm-5.2, kimi-k2.6, minimax-m2.7; DeepSeek V4 series (deepseek-v4-flash-0731, deepseek-v4-pro-0813) launched on 2026-08-18
• 2026/08/27: Alibaba Cloud Bailian Token Plan (Personal Edition) added model qwen3.8-flash
• 2026/08/27: Volcano Engine Coding Plan (Personal Edition) and Agent Plan DeepSeek-V4-Pro changed from "early-access preview" to officially released (significantly enhanced Agent capabilities, accessible via model name and console selection)
• 2026/08/27: Alibaba Cloud Bailian model pricing adjustments: qwen3.8-flash China input dropped from 1 to 0.8 RMB and output from 3 to 2.7 RMB (international input dropped from 1.167 to 1.094 RMB, output unchanged); added kimi-k3 (thinking-only mode, global 20/100 RMB, international 21.875/109.376 RMB per 1M tokens); tongyi-xiaomi-analysis-flash/pro now support context cache discounts; added kling/kling-v3-turbo video generation model (audio video 720P 0.8 RMB/sec, 1080P 1.0 RMB/sec)
• 2026/08/27: Alibaba Cloud Bailian context cache rule adjustments: implicit cache minimum token count raised from 256 to 1,024 (same as explicit cache, but meaning differs — reaching 1,024 only qualifies for a hit, does not guarantee one); qwen3.8-flash added to cached_token discount exceptions (no longer billed at 20% of input); added tongyi-xiaomi-analysis-pro/flash industry models; added DeepSeek (Kuaishou Wanqing deployment) cache pricing (vanchin/deepseek-v4-pro 8.33%, vanchin/deepseek-v3.2-think 10%, etc.)
• 2026/08/27: Tencent Cloud Token Plan (Personal Edition) Universal Plan Kimi-K2.5 discontinuation date changed from 2026-07-31 to 2026-08-31 (and removed the peak-hour rate-limit notice)
• 2026/08/26: Zhipu added GLM-5.3-Flash API pricing (bigmodel.cn): cached input 0.115/0.23 RMB, uncached input 0.4/0.8 RMB, output 1.4/2.8 RMB per 1M tokens, 1M context, currently 50% off for two weeks
• 2026/08/26: Zhipu GLM Coding Plan supported models changed from {GLM-5.3, GLM-5-Turbo, GLM-4.7} to {GLM-5.3, GLM-5.3-Flash}; GLM-5.3-Flash (320B total params / 18B active, hybrid linear+sparse attention, native vision) launched for Coding Plan with deduction coefficients Input 2.3 / Cached Input 0.56 / Output 8 (including visual understanding MCP); GLM-5-Turbo/GLM-4.7 requests auto-routed to GLM-5.3-Flash; token allowance reference table restructured to show per-model breakdowns at 95%/96%/98% cache hit rates; "save up to 92%" claim restored
• 2026/08/26: OpenCode Go added GLM-5.3-Flash model, removed Ox Alpha Free model (limited-time free ended)
• 2026/08/26: Qwen3.8-Flash-Next and GLM-5.3-Flash released
• 2026/08/26: MiniMax Token Plan price adjustment: International edition Plus/Max/Ultra plans increased from \(20/\)50/$120/month to \(22/\)55/$132/month; Chinese edition prepaid credits packages adjusted — ¥30 for 4,489 credits (was 4,285), ¥150 for 22,460 credits (was 21,430), ¥500 for 74,900 credits (was 71,435); Chinese subscription plan prices unchanged
• 2026/08/26: OpenCode Go upgraded Grok model from 4.5 to 4.6: rate limits increased (169/5hr, 423/week, 845/month, previously 120/300/600), pricing changed to tiered (≤200K tokens: input $2.00, output $6.00, cache $0.50; >200K tokens: input $4.00, output $12.00, cache $1.00, previously flat \(2.00/\)6.00/$0.30), model ID changed from grok-4.5 to grok-4.6
• 2026/08/25: Tencent Cloud Token Plan Enterprise Professional Plan lowered minimum purchase from 100K to 50K Credits, added DeepSeek-V4-Flash 0731 Official and DeepSeek-V4-Pro 0813 Official models (peak-valley pricing, off-peak rates same as direct-supply versions)
• 2026/08/25: OpenCode Go added support for LongCat-2.0 model (input $0.30/1M, output $1.20/1M, cache read $0.006/1M)
• 2026/08/24: OpenCode Go removed first-month discount, pricing changed from "$5 first month, then $10/month" to a flat $10/month
• 2026/08/23: DeepSeek API peak-valley weekend rule now in effect: peak hours are explicitly Monday to Friday 9:00-12:00, 14:00-18:00 Beijing time; weekends are now charged at off-peak rates all day
• 2026/08/22: DeepSeek API peak-valley pricing rule adjustment: starting 2026-08-23 (Sunday) 00:00 Beijing time, weekends (Saturday and Sunday) will no longer distinguish peak/off-peak — all day will be charged at off-peak rates (previously weekends still followed peak/off-peak schedule)
• 2026/08/21: Volcano Engine Coding Plan (Personal Edition) added support for Doubao-Seed-Evolving model (for Coding & Agent scenarios, weekly upgrades, 1M context window, 256K max output); OpenCode Go added support for DeepSeek V4 Flash Vision Exp model (pricing Off-Peak \(0.22/\)0.66, Peak \(0.44/\)1.32 per 1M tokens, images converted to tokens based on size); Tencent Cloud Token Plan (Personal Edition) Universal Plan removed discontinued models (Tencent HY 2.0 Instruct, Tencent HY 2.0 Think, Hunyuan-T1, Hunyuan-TurboS, MiniMax-M2.5)
• 2026/08/21: DeepSeek added new vision model deepseek-v4-flash-vision-exp (experimental), priced same as deepseek-v4-flash, FIM completion not supported, concurrency limit 2500, images converted to tokens based on size; Tencent Cloud Token Plan Enterprise Professional Plan removed MiniMax-M2.5 model; Volcano Engine Agent Plan (Personal Edition) updated quota rules: image/video generation models, voice models, and Harness merged into a single daily quota category (no longer separated as "vision models" and "voice models"), daily quota uniformly set to half of monthly plan quota; OpenCode Go added Ox Alpha Free model (limited-time free)
• 2026/08/20: CTCloud Programming Token Plan updated supported models: added GLM-5.1, DeepSeek-V4-Flash-0731; renamed GLM-5 to GLM-5.0 (Official); DeepSeek-V3.2 labeled as Flagship
• 2026/08/19: Alibaba Cloud Bailian added open-source model qwen3.8-27b, priced at Input 3 RMB / Output 12 RMB per 1M tokens (context cache discount available), international pricing Input 3.646 RMB / Output 21.875 RMB per 1M tokens; Tencent Cloud Token Plan (Personal Edition) and Enterprise Professional Plan renamed DeepSeek-V4-Pro 原厂直供 to DeepSeek-V4-Pro 正式版 原厂直供, added model IDs deepseek/deepseek-v4-pro-0813 and deepseek/deepseek-v4-pro; MiniMax Token Plan coverage narrowed from "all models" to "flagship models", music APIs (Music-3.0, Music-2.6, Lyrics Generation, etc.) discontinued and removed from Token Plan quota
• 2026/08/19: Zhipu released GLM-5.3 model with 50% coding improvement over GLM-5.2 and cybersecurity capabilities matching Mythos 5; GLM Coding Plan token allowance estimates updated to show multiple cache hit rates (90.9%, 95%, 98%); GLM-5.3 API pricing matches GLM-5.2
• 2026/08/18: Volcano Engine Coding Plan (Personal Edition) and Agent Plan removed MiniMax-M2.7 and Kimi-K2.6 models (previously marked as phasing out)
• 2026/08/17: Volcano Engine Coding Plan (Personal Edition) and Agent Plan removed Doubao-Seed-2.0-Code, Doubao-Seed-2.0-pro, Doubao-Seed-Code models (previously marked as phasing out); GLM-5.2 marked as phasing out; GLM-5.3 replaced GLM-5.2 as the default glm-latest target
• 2026/08/14: Volcano Engine Coding Plan (Personal Edition) and Agent Plan added support for the GLM-5.3 model (1M context window, 1024K context / 128K max output, thinking enabled by default and cannot be disabled, deduction coefficient same as GLM-5.2)
• 2026/08/14: Zhipu GLM Coding Plan flagship model upgraded from GLM-5.2 to GLM-5.3: all plans support GLM-5.3, GLM-5-Turbo and GLM-4.7; requests for previous models (GLM-5.2/GLM-5.1) are automatically routed to GLM-5.3; model deduction coefficients unchanged (Input 6.9 / Cached Input 1.7 / Output 24); the official docs also removed the "save up to 92%" claim; the international edition (Z.ai DevPack) was upgraded to GLM-5.3 accordingly
• 2026/08/13: DeepSeek API will adopt peak-valley pricing starting 2026-08-17 00:00: peak hours (9:00-12:00, 14:00-18:00 Beijing time) are charged at 2x the off-peak price, with overall prices significantly increased (e.g., deepseek-v4-pro output price rises from 6 RMB to 27 RMB peak / 13.5 RMB off-peak per 1M tokens)
• 2026/07/31: GLM Coding Plan switched to credit-based limits
• 2026/07/20: Kimi Code separated from Kimi membership, now available as a dedicated new Kimi Code subscription
• 2026/07/16: Kimi-K3 model released
• 2026/07/13: Baidu Qianfan Token Plan (Personal Edition) launched
• 2026/06/27: Infini Coding Plan discontinued
• 2026/06/13: GLM-5.2 model released
• 2026/06/12: Kimi K2.7-Code model released
• 2026/06/10: GLM Coding Plan Team Edition launched
• 2026/06/08: Volcano Engine Coding Plan and Agent Plan added MiniMax-M3
• 2026/06/08: MiniMax M3 API price permanently halved
• 2026/06/07: Alibaba Cloud Bailian Coding Plan added qwen3.7-plus model
• 2026/06/05: Huawei Cloud MaaS Token Plan launched
• 2026/06/01: MiniMax M3 released
• 2026/05/29: StepFun Coding Plan added support for step-3.7-flash model
• 2026/05/27: Xiaomi MiMo Token Plan quota massively increased
• 2026/05/22: Baidu Qianfan Coding Plan added support for DeepSeek-V4-Pro model
• 2026/05/22: Alibaba Cloud Token Plan added support for qwen3.7-max model
• 2026/05/08: Baidu Qianfan Coding Plan added support for DeepSeek-V4-Flash and GLM-5.1 models
• 2026/05/07: Volcano Engine Agent Plan (Personal Edition) launched
• 2026/04/30: Tencent Cloud Token Plan (Personal Edition) added support for GLM-5.1 and MiniMax-M2.7 models
• 2026/04/30: StepFun Coding Plan removed deepseek-v4-pro model, must access indirectly through step-router-v1 model
• 2026/04/28: StepFun Coding Plan added deepseek-v4-pro and step-router-v1 models
• 2026/04/27: GLM Coding Plan extended the limited-time discount benefit deadline from end of April to end of June
• 2026/04/24: Alibaba Cloud Bailian Token Plan halved the Credits required for cached input tokens
• 2026/04/23: MiMo-V2.5 series models launched
• 2026/04/23: StepFun Coding Plan added support for stepaudio-2.5-asr model
• 2026/04/23: GLM Coding Plan will uniformly disable auto-renewal for legacy plans (no weekly limit version) on April 30, 2026. Current active billing cycles are not affected. Meanwhile, the system will automatically gift affected users 2 months of equivalent new plan, which will take effect after the current plan expires, with no manual claim required. See Legacy Plan Migration and Compensation Notice.
• 2026/04/22: Volcano Engine Coding Plan added MiniMax-M2.7, Kimi-K2.6, GLM-5.1
• 2026/04/21: Alibaba Cloud Bailian Token Plan (Team Edition) launched
• 2026/04/21: Kimi officially released Kimi-K2.6 model
• 2026/04/14: Kimi Code launched K2.6-code-preview model
• 2026/04/12: Zhipu International GLM Coding Plan starting price increased from 10 USD/month to 18 USD/month
• 2026/04/11: Alibaba Cloud Bailian Coding Plan Lite Plan will stop accepting renewals and upgrades from April 13, 2026. New purchases were already suspended on March 19, 2026
• 2026/04/11: Added CTCloud Coding Plan
• 2026/04/09: Infini-AI Coding Plan added support for glm-5.1 model
• 2026/04/09: Zhipu Coding Plan removed GLM-5, GLM-4.6, GLM-4.5 models
• 2026/04/08: iFlytek Astron Coding Plan launched new Updated Edition plans; old First Month Edition plans retired
• 2026/04/08: Alibaba Cloud Bailian Coding Plan added recommended model qwen3.6-plus (with image understanding), available only on the Pro plan. qwen3.5-plus was moved from recommended to more models
• 2026/04/07: Baidu Qianfan Coding Plan removed GLM-4.7 and MiniMax-M2.1, added ERNIE-4.5-Turbo-20260402
• 2026/04/03: Added Xiaomi MiMo Token Plan
• 2026/04/03: Added JD Cloud Coding Plan
• 2026/04/03: Added StepFun Coding Plan
• 2026/03/27: GLM-5.1 launched in GLM Coding Plan
• 2026/03/27: Added Tencent Cloud LLM Token Plan, which uses token-based limits instead of request counts compared to the Coding Plan
• 2026/03/26: GLM-5-Turbo is now available on all GLM Coding Plan tiers; previously only available on Max
• 2026/03/21: MiniMax Token Plan has brought back the Starter Plan, with the same pricing and limits. Additionally, a weekly limit has been introduced, which is 10 times the limit per 5 hours.
• 2026/03/19: Alibaba Cloud Bailian Coding Plan Release Announcement, new purchases of the Coding Plan Lite will be suspended from 2026-03-20 00:00:00 Beijing Time.
• 2026/03/19: Infini-AI Coding Plan added support for third-party model minimax-m2.7
• 2026/03/18: MiniMax Token Plan removed the statement that the MiniMax-M2.7-highspeed version consumes double the requests.
• 2026/03/18: MiniMax-M2.7 launched; meanwhile, MiniMax Coding Plan was renamed to MiniMax Token Plan, supporting non-text LLMs (e.g., audio and video). The Token Plan removed the Starter Plan and changed the description from "prompts" to "requests" though the actual limits remain unchanged (previously, the limit was also calculated as 1 prompt equaling 15 requests)
• 2026/03/17: Added iFlytek MaaS Astron Coding Plan
• 2026/03/16: Zhipu launched the GLM-5-Turbo model, described as follows:
  • A base model deeply optimised for the OpenClaw lobster scenario
  • Strengthened calling of external tools and various Skills, more stable and reliable in multi-step tasks
  • Stronger decomposition of complex instructions — it identifies goals precisely, plans steps, and supports coordination and division of labour across multiple agents
  • Better understanding of temporal requirements, maintaining execution continuity in long, complex tasks
  • Further improved execution efficiency and response stability for lobster tasks with high data throughput and long logic chains
  • GLM-5-Turbo plan availability: already supported on Max; expected on Pro by the end of March and on Lite within April
  • GLM-5 plan availability: already supported on both Max and Pro; expected on Lite by the end of March
  • GLM-5 and GLM-5-Turbo are advanced models benchmarked against Claude Opus, consuming quota at "3x during peak hours, 2x during off-peak hours"; we recommend switching to GLM-5 for complex tasks and continuing with GLM-4.7 for ordinary tasks to avoid burning through the plan quota too quickly. (As a limited-time benefit, GLM-5-Turbo consumes only 1x during off-peak hours, through the end of April.) Note: peak hours are 14:00–18:00 (UTC+8) daily
• 2026/03/08: Tencent Cloud LLM Coding Plan launched
• 2026/03/07: Zhipu issued a 15-day compensation credit for GLM Coding Plan. The translated email text is as follows (original was in Chinese):

  Dear GLM Coding Plan User,
  
  
  Thank you for your continued support and trust.
  
  
  To address the recent service issues some users experienced and to express our sincere apologies, we have credited your account with compensation equivalent to 15 days of your current subscription fee (with no expiration). You can view the credit details in your account dashboard under "Zhipu Open Platform Console - Billing - Transaction History", and it will be automatically applied to your future usage.
  
  
  Thank you again for your understanding and patience, as well as your ongoing support. We remain committed to improving our product capabilities and service quality to deliver a more stable and efficient development experience.
  
  
  Best regards,
  
  
  Zhipu AI Platform Team
  
  March 7, 2026
  

• 2026/03/06: Volcano Engine Coding Plan added support for third-party model MiniMax-M2.5
• 2026/02/25: Alibaba Bailian Coding Plan added support for third-party model minimax-m2.5
• 2026/02/24: Alibaba Bailian Coding Plan added support for third-party model glm-5
• 2026/02/21: Observed that Alibaba Bailian Coding Plan added support for third-party models glm-4.7 and kimi-k2.5; previously only Qwen's own models were available
• 2026/02/18: Kimi Code billing has changed:
  • Previously, the weekly limit was changed from 50M input + output tokens to 4M uncached input + output tokens, while the per-5-hour limit remained 10M input + output tokens
  • Now the per-5-hour limit has been changed to 1M uncached input + output tokens
  • Therefore, there is now a 4x relationship between the per-5-hour limit and the weekly limit
  • Assuming 99.5% input (95% cached, 5% uncached) + 0.5% output ratio, old vs new limit comparison:
    • Old weekly limit 50M input + output tokens: 50M*0.5%=250K output tokens
    • New weekly limit 4M uncached input + output tokens: 4M*0.5%/(0.5%+99.5%*5%)=365K output tokens
    • Old per-5-hour limit 10M input + output tokens: 10M*0.5%=50K output tokens
    • New per-5-hour limit 1M uncached input + output tokens: 1M*0.5%/(0.5%+99.5%*5%)=91K output tokens
  • Assuming 99.5% input (90% cached, 10% uncached) + 0.5% output ratio, old vs new limit comparison:
    • Old weekly limit 50M input + output tokens: 50M*0.5%=250K output tokens
    • New weekly limit 4M uncached input + output tokens: 4M*0.5%/(0.5%+99.5%*10%)=191K output tokens
    • Old per-5-hour limit 10M input + output tokens: 10M*0.5%=50K output tokens
    • New per-5-hour limit 1M uncached input + output tokens: 1M*0.5%/(0.5%+99.5%*10%)=48K output tokens
  • As can be seen, whether the new or old limit is more restrictive depends on the cache hit rate
• 2026/02/16: GLM Coding Plan increased weekly limits from 4x the per-5-hour limit (320/1600/6400 prompts) to 5x (400/2000/8000 prompts). Meanwhile, GLM-5 consumption rate changed from 3x to 3x during peak hours and 2x during off-peak hours (peak hours: 14:00-18:00 UTC+8 daily).
• 2026/02/16: Recently discovered some changes in Kimi Code billing:
  • The Andante plan's per-5-hour limit remains unchanged at 10M input + output tokens, but the weekly limit observed when opening a new Code Session depletes faster, clearly not being 20% of the per-5-hour usage (previous calculation showed weekly limit was 5x the per-5-hour limit). However, with continued use, the ratio stays around 20%. Calculated using the previous method, weekly usage is approximately 48M input + output tokens rather than the original 50M—a rather odd number.
  • This question was explained by LLM Inference Systems, Code Agents, and Power Grids - Xu Xinran: cached tokens are not counted in usage.
  • If calculated by uncached input + output tokens, then weekly usage is 4M uncached input + output tokens; the 5-hour limit should still use the old calculation of 10M input + output tokens.
  • The purpose is that if Kimi Code is used for non-Vibe Coding scenarios with low cache hit rates, the weekly limit will be consumed quickly.
  • Further reading: suspiciously precise floats, or, how I got Claude's real limits
• 2026/02/15: MiniMax Coding Plan added Plus/Max/Ultra High Speed versions
• 2026/02/14: GLM Coding Plan added weekly limits at 4x the per-5-hour limit (Kimi is 5x, Volcano and Alibaba are 7.5x). Meanwhile, GLM-5 consumes quota at 3x the rate of GLM-4.7.
  • Unserious comment: Looks like at Zhipu, you only work 4 days a week, 5 hours a day, while at Moonshot you work 5 days a week, and at ByteDance and Alibaba you work 7.5 days a week. Which company has more overtime is clear at a glance, lol. (But ByteDance and Alibaba only work for two weeks a month; the other two weeks are off, this is the "big week small weeks" system?)
  • Serious comment: The new GLM Coding Plan's cost-effectiveness dropped from "夯"(S-level) to "NPC"(C-level), making Kimi/MiniMax's value stand out. The solution is to continue renewing the old plan and stick with GLM-4.7.
  • If calculated using the new plan at 2/3 of the old quota limit, using GLM-4.7, the Lite plan can use 40M*2/3*4*30/7=457M tokens per month (30 days); using GLM-5, it would be 40M*2/3*4*30/7/3=152M tokens.
• 2026/02/12: GLM Coding Plan prices changed from 40/200/400 RMB/month to 49/149/469 RMB/month; meanwhile, usage quotas were reduced to 2/3 of the original:
  • Lite Plan: Maximum ~80 prompts per 5 hours (was 120), equivalent to 3x Claude Pro plan usage
  • Pro Plan: Maximum ~400 prompts per 5 hours (was 600), equivalent to 5x Lite plan usage
  • Max Plan: Maximum ~1600 prompts per 5 hours (was 2400), equivalent to 4x Pro plan usage
  • If using the 2/3 ratio, the Lite plan limit would be 40/3*2=27M tokens per 5 hours. The new version also has weekly limits (specific rules published on 2026/02/14, see above). After switching to the new plan (I am not planning to), need to test what the new usage limits correspond to in tokens (interested readers can test and provide feedback).
• 2026/02/12: Added description for Kimi Allegro plan
• 2026/02/12: With the release of GLM-5, GLM Coding Plan's quota/limit API no longer returns specific token counts—presumably preparing for GLM-5 and GLM-4.7 to consume usage at different rates (based on API pricing, guessing there might be a 2x coefficient? Awaiting further testing). However, current testing shows GLM-4.7 usage limits remain unchanged; Lite plan is still 40M input + output tokens per 5 hours. Since there's only a per-5-hour limit, calculated at 30 days per month, theoretically maximum monthly usage could be 30*24/5*40=5760M tokens.
• 2026/01/30: Through actual testing, speculated that GLM Coding Plan's Lite plan usage limit is that the sum of all requests' input + output tokens does not exceed 40M per 5 hours (meaning each prompt corresponds to 40M/120=333K tokens), which is consistent with the results returned by the https://open.bigmodel.cn/api/monitor/usage/quota/limit API (after 2026/02/12, this API only returns percentages, not token counts).

Cite This

Comments