AI Deals & Discounts
Discover the best AI subscriptions and API pricing deals
TokenRhythm 域名迁移确认 — tokenrhythm.com 已过期，真实官网为 tokenrhythm.studio
As of October 23, 2026, tokenrhythm.com returns HTTP 302 to a Dynadot domain-for-sale page, confirming the domain has lapsed. The live official properties are tokenrhythm.studio (console and referral links, HTTP 200) and tokenrhythm.ai (brand site, HTTP 200), while tokenrhythm.dev and docs.tokenrhythm.dev no longer resolve. TokenRhythm continues to offer a CNY 68 sign-up credit, a CNY 1 per million input and output aggregator rate across 16 models, and a 1M context tier. All outbound links on aicup.vip have been repointed to the working domain.
MiniMax 双线并行最低价 — M3.1-Flash ¥1/¥4 + M3 ≤512K「永久五折」¥2.10/¥8.40
MiniMax now runs two parallel floors. M3.1-Flash is priced at CNY 1.00 input / CNY 4.00 output per million tokens with CNY 0.10 cached reads, while M3's 50% discount below 512K input tokens is labelled permanent rather than promotional, holding at CNY 2.10 / CNY 8.40 with CNY 0.42 cache. Above 512K, M3 doubles to CNY 4.20 / CNY 16.80. Priority service via service_tier=priority is 1.5x the standard rate.
GLM-5.3-Flash 限时五折 ¥0.40/¥1.40 — 10/28 到期，剩 5 天（缓存 ¥0.115/M）
Zhipu's limited-time 50% discount on GLM-5.3-Flash expires on October 28, 2026, cutting the price to CNY 0.40 input / CNY 1.40 output per million tokens with CNY 0.115 cached input. The standard rate is CNY 0.8/2.8. Zhipu positions the Flash tier as matching Claude Opus 4.8 on selected capabilities at roughly 1/40 of the cost, with a 1M context window. Batch API adds a further 50% off. GLM-4.7-Flash and GLM-4.6V-Flash remain entirely free.
Claude Managed Agents 上线 $0.08/session-hour — 空闲不计费，长任务智能体成本可预测
Anthropic's managed agents bill session runtime at $0.08 per session-hour, metered by the millisecond but only while a session is in the running state; idle, rescheduling and terminated sessions are not charged. Token consumption matches the underlying model and still honours prompt caching, Fast Mode and data residency modifiers. Batch API discounts and CCU billing do not apply. Official example: a one-hour Opus 5.5 session using 50K input and 15K output tokens costs $0.58, or $0.428 if 40K of the input is cached.
Claude Web Fetch 完全免费 + Code Execution 每月 1,550 小时免费（首次完整披露）
Anthropic's pricing page now documents tool-level billing in full. Web Search costs $10 per 1,000 successful searches and is not charged on failure. Web Fetch is free and only consumes standard token charges, with max_content_tokens available to cap the payload. Code Execution is free when used alongside Web Search/Fetch, or $0.05 per container hour when used standalone, with 1,550 free hours per organization per month and a five-minute minimum. Bash tool on 4.7+ models adds 325 input tokens; text editor adds 974.
Grok 4.3 拿下 xAI 唯一 1M 上下文档 — $1.25/$2.50，比 Grok 4.7 便宜 37%
xAI's model table confirms grok-4.3 at $1.25 input / $2.50 output per million tokens below 200K input, jumping to $2.50/$5.00 for the whole request at or above 200K, with $0.20 cached input. Unlike grok-4.5, 4.6 and 4.7 which all sit at $2/$6 but cap context at 500K, grok-4.3 carries a 1M context window. grok-4.20 reasoning / non-reasoning / multi-agent share the same tier.
讯飞星火 X2.6 发布 — 输入 ¥0.9/M + 缓存 ¥0.08/M + 1M 上下文，4B 蒸馏版同步开源
iFlytek released Spark X2.6 on October 22, 2026 at the 1024 Developer Festival: a 293B-total / 30B-active end-cloud unified MoE priced at CNY 0.9 per million input tokens, CNY 0.08 per million cached input tokens and CNY 3.6 per million output tokens, with a 1M context window. A 4B distilled version was open-sourced at the same time, making it one of the few 1M-context Chinese models available both as an API and as open weights.
OpenAI Decisions API 正式 GA — $0.10/M 输入且输出不单独计费，单请求 400 问（10/22）
OpenAI moved the Decisions API from research preview to general availability on October 22, 2026. Built on GPT-6 Luna, it charges $0.10 per million input tokens and $0.01 per million cached input tokens, and does not bill output tokens separately. A single request accepts up to 400 questions (twice the previous limit) and runs up to 10x faster than GPT-6 Luna through the Responses API. For batch ranking, triage and classification workloads the unbilled output side makes the effective per-question cost lower than any output-billed model.
Anthropic 开放 METR 员工级访问 — 独立评估员免费研究权限（10/15）
Anthropic granted METR 6 employees hardware/office/internal system access with publication rights not subject to Anthropic veto. SecureBio reviewed bio/chem chapter. Accenture/Faculty/Hugging Face at varying cooperation phases. Sets precedent for independent evaluator model access.
OpenRouter 24 :free 模型扩展 — 增加 DeepSeek V4-Pro 1M / GLM-5.3-Flash 1M / Kimi K3（10/16 上线）
OpenRouter expanded :free tier to 24 models 10/16: added DeepSeek V4-Pro 1M context, GLM-5.3-Flash 1M, Kimi K3, plus the previous 21 models. :free models rate-limited to 1 req/s but unlimited token volume. $10 sign-up credit still active.
Apple Intelligence 中国 12 月上线 iOS 27.3 — 免费档、不含最新 Siri AI、本地合作伙伴合规通道
Apple Intelligence China launches with iOS 27.3 in December 2026. Free for iPhone 15 Pro / iPhone 16 / iPhone 17 series and M1+ iPad owners. Initial launch excludes latest Siri AI; uses compliant routing through Alibaba/Baidu/Baichuan partner channels.
阿里「千问 Go」个人代理免费档 — 免登录跨端 + 试穿试戴 + 行程规划
Alibaba launched Qwen Go personal agent 10/15 as a free tier across web/iOS/Android. No login required for casual use. Built on Qwen3.8-Max with shopping agent, virtual try-on, travel planning, and Qwen App sync. Free through 12/31.
Anthropic Claude Marketplace — 19 合作伙伴产品 + 首月承诺预算抵扣
Anthropic Claude Marketplace launched 10/14 with 19 initial partner products (Slack/Notion/Linear/Box/Snowflake/etc.). Use committed Claude budgets to purchase Marketplace add-ons; first-month waiver for new commerce customers through 11/30.
xAI Grok Voice 3.0 首发 $0.08/min 试用 — 最低语音模型定价 + SpaceX/Tesla 直连
Grok Voice 3.0 launch trial: $0.08/min pricing (lowest among major voice models, 5-10x cheaper than GPT-4o Realtime equivalent). Built-in Starlink customer service and Tesla in-car Grok 4 Voice support. Trial ends when Grok Voice 3.0 hits GA in November.
OpenAI Rosalind 学术计划 — 认证研究人员 GPT-6 API 额度 + 算力 + 研究访问
OpenAI Rosalind is an academic access program providing certified university researchers with GPT-6 series API credits, research preview access, and dedicated compute hours. Apply at openai.com/rosalind with institutional email + advisor letter.
Gemini 4 Argon Global GA — API 50% Off Through 11/30 ($2/$10 per 1M)
Gemini 4 Argon launched globally Oct 14 with 50% intro pricing: $2/$10 per 1M tokens, cached input 95% off ($0.10/M). 1M context + 1M single-shot output. AA Intelligence Index 53. Pricing reverts to $4/$20 from Dec 1.
Claude Marketplace 19 Partners — Use Existing Commitment, Zero Commission
Anthropic launched Claude Marketplace Oct 14 with 19 Claude-powered partner products: Code (Augment, Cursor, Lovable, Vercel, GitLab), Security (CrowdStrike Charlotte AI, Cyera, Chainguard, Checkmarx), Legal (Harvey, Legora), Finance (Hebbia, Rogo), Data (Gamma, Snowflake). Eligible enterprises route portion of existing Anthropic commitment to buy partner tools, zero commission (vs AWS/Azure 3-15%).
Grok Voice 3.0 Launches — $0.08/min, Speech-to-Speech Index #1 (86.9)
xAI launched Grok Voice 3.0 Oct 14 at $0.08/min (unchanged from 2.0). Speech-to-Speech Index 86.9 #1 (vs GPT-Live-1 Astra 83.2, Gemini 3.8 Live Extended Thinking 82.6). First audio response 0.81s. Powers Tesla vehicle voice + Starlink call center (20K calls/day, 70% auto-resolved).
Qwen Go! Personal Agent — World's First End-to-End Personal Agent
Alibaba launched Qwen Go! Personal Agent in 千问 App Oct 13: world's first end-to-end personal AI agent. Capabilities: flight delay refund negotiation (got user 70元 cash), Taobao price-protection one-click, Tuhu car maintenance booking, badminton court booking, cross-App purchase comparison. Powered by Qwen 3.8 series Agentic.
MiniMax Tmall Flagship — M Plan Go/Explore/Build, First Order 5x Off 24.5元/月
MiniMax opened first official Tmall flagship store Oct 13 ahead of 双11. M Plan offers Go/Explore/Build subscriptions (monthly + annual). First order 5x off through 10/14, starting 24.5元/月. Joins Zhipu/Aliyun/Tongyi Office on Tmall 'AI Hub' Token recharge center.
GPT-6 Intelligent UI Pushed to Free ChatGPT (10/14) — GPT-5.5 Series Retired
OpenAI fully pushed GPT-6 + Intelligent UI to free ChatGPT on Oct 14. Free/Go get GPT-6 Luna, Plus/Business/Enterprise get GPT-6 Sol, Pro gets GPT-6 Sol/Astra. GPT-5.5 Luna/Sol/Terra officially retired. Intelligent UI uses native streamable components + compiler to generate charts/forms/calculators mid-reply. ChatGPT weekly active users now 1.2B.
Coding Plan 3-Tier Stack (Wed 10/14 W42) — Cursor Pro + Kimi Code + GLM Coding Max
Coding Plan 3-Tier Stack: Cursor Pro $16/月（IDE 体验之王） + Kimi Code Andante ¥49/月（国产性价比）+ GLM Coding Max ¥120/月（2000 次 GLM-5.3-Flash + 50 次 Claude Sonnet 5）。合计 $16 + ¥169 = ¥282/月 ≈ $39，按 IDE 体验 / 国产性价比 / IDE 模型全覆盖 三档分层。
Aliyun Token Plan Lite ¥39/月 (vs 原价¥60) — Token Plan 入门首选
阿里云百炼 Token Plan 个人版 Lite：¥39/月（原价¥60），含 700 Credits/5h、2,500 Credits/7天，支持 1-2 个 Agent 并发，可抢先体验 Qwen3.8-Max-Preview。Qwen3.8-Max-Preview 限时白天 Credits 消耗 1 折、夜间 22:00-08:00 再享 2 折折上折 (20x)。仅限编程工具和智能体工具使用（Claude Code/OpenClaw/OpenCode/Cursor/Qwen Code等）。
TokenRhythm 1M Aggregator ¥1/¥1 per M (¥500 credit 延期至 10/31) — 跨厂商聚合
TokenRhythm API 聚合网关 (https://tokenrhythm.studio/) 提供 5 款 1M 上下文模型统一入口：DeepSeek V4-Pro ¥1/¥1、Qwen3.8-Max ¥1/¥1、Kimi K3 ¥1/¥1、GLM-5.2 ¥1/¥1、MiniMax-M3 ¥1/¥1。新用户 ¥500 体验金延期至 10/31，100M token 跑批 ¥200。
Cursor Pro Annual — $192/yr ($16/mo, save $48 vs monthly)
Cursor Pro annual plan at $192/year ($16/mo equivalent) saves $48/year vs monthly $20/mo = $240/year. Includes 500 fast premium requests/mo, GPT-5.6 + Claude Sonnet 5 + Gemini 3.7 Flash access, unlimited slow requests, Background Agent, Bug Bot. Best value for solo developers who already use Cursor >6mo. Available only via direct annual subscription.
TokenRhythm 1M Gateway — All-in-One API Aggregator ¥1/¥1 per M
TokenRhythm API gateway (https://tokenrhythm.studio/) provides unified access to 11+ vendors (OpenAI/Anthropic/Google/DeepSeek/GLM/Qwen/Kimi/MiniMax/Mimo/xAI/TokenRhythm native) with 1M context + multi-region failover + international card payment. 1M-context pricing ¥1 input / ¥1 output per M tokens — competitive vs Gemini 3.7 Flash 1M ($0.30/$1.20 ≈ ¥2.2/¥8.7) and way below Claude Opus 4.8 1M ($15/$75 ≈ ¥110/¥550). Saves $30-200/mo vs maintaining 5+ vendor accounts separately.
Coding Plan 3-Tier Stack (Wed 9/23) — Cursor Pro + Kimi Code + GLM Coding Max
Curated Coding Plan 3-tier stack for Q3-Q4 2026: (1) <$15/mo tier — Kimi Code Allegretto ¥99 (Chinese-first AI coding, 256K context); (2) $15-30 tier — Cursor Pro $20/mo or GLM Coding Max ¥120/mo (both 500 fast requests + best IDE integration); (3) $30+ tier — Cursor Business $40 or GitHub Copilot Business $39 (team/enterprise features). Total for solo developer: <$60/mo for full coverage.
Qwen3.8-Flash 1M Context — Cheapest Domestic 1M LLM
Qwen3.8-Flash supports 1M token context at ¥1.2 input / ¥4 output per M tokens — the cheapest domestic 1M-context API as of Tue 9/22/2026. Compared to international peers: Gemini 3.7 Flash 1M $0.30/$1.20 (≈¥2.2/¥8.7); Claude Opus 4.8 1M $15/$75 (≈¥110/¥550); GPT-5.6 Sol 1M ($5+$5)/($30+$30) — so Qwen3.8-Flash 1M is 60%+ cheaper than Gemini and 92%+ cheaper than Claude for long-context workloads.
Qwen 3.8 Flash 1M — 长文 RAG 性价比第一 ¥1.2 input / ¥4 output per M (1M context)
Alibaba Qwen 3.8 Flash now supports 1M context window at ¥1.2 input / ¥4 output per M tokens. W38 best value for 1M RAG workloads: code repo Q&A, legal contract review, technical doc search.
GPT-5.6 Luna 海外最低 — $0.20 input / $1.20 output per M (W38 overseas cheapest)
OpenAI GPT-5.6 Luna remains at -20%/-80% pricing through W38: $0.20 input / $1.20 output per M tokens. The cheapest overseas flagship-tier API. Suitable for translation, classification, and large-scale batch workloads.
Full Multimodal 3-Stack Combo — MiniMax Speech 2.8 HD + Doubao Seedream 4.0 + Sora 2 = $70/mo
Three-stack full multimodal combo at $70/month covers text/image/audio/video: (1) MiniMax Speech 2.8 HD (TTS premium, $30/mo for 1000 minutes); (2) Doubao Seedream 4.0 (image generation, $20/mo for 2000 images); (3) Sora 2 (video generation, $20/mo for 100 clips). All layers via official channels, no aggregator. Best for: content creators, marketing teams, video production studios.
Domestic Productivity 3-Stack Combo — Bailian Qwen3.8-Flash + Volcano Doubao + tokenrhythm = $68/mo for 50M tokens
Three-stack combo for Chinese productivity at $68/month for 50M mixed tokens: (1) Alibaba Bailian Qwen3.8-Flash at ¥0.4/M (1M context, Chinese-optimized, ¥200 credits for new accounts); (2) Volcano Engine doubao-2.0-pro ¥100 trial credit (covers ~5M multimodal tokens); (3) tokenrhythm ¥500 credit + DeepSeek V4.1 Flash valley pricing ¥1/M (23:00-08:00). Total: ~50M tokens for $68 vs $300+ using OpenAI directly. Best for: Chinese teams, content creators, cross-modal workflows.
Claude Max 5x Annual — Sunday combo $100/yr (5x Pro usage, best dev value)
Claude Max 5x is Anthropic's annual plan tier giving 5x Pro usage for $100/year (vs Pro $20/month × 12 = $240/yr). Effective cost $8.3/month with 5x Pro limits. Unlocks: Claude Sonnet 5, Opus 4.8, Haiku 4.5, Projects, Artifacts, custom styles. Best for: heavy Claude users doing coding + writing + research.
ChatGPT Plus Annual — Sunday combo save $40/yr ($200/yr vs $20/mo)
ChatGPT Plus annual subscription saves $40/year vs monthly billing ($200/yr vs $240/yr if paid monthly). Unlocks: GPT-5.6 Sol/Terra/Luna (200 messages/week), GPT-5.6 Pro ($200/yr Pro tier), image generation (DALL-E 4), file uploads, advanced data analysis, custom GPTs. 30-day refundable.
Cursor Pro Annual Plan — Sunday combo $192/yr (same as monthly, no annual discount yet)
Cursor Pro annual billing at $192/year equals monthly billing $20×12=$240/year — currently no annual discount. However, Pro plan unlocks: 500 fast premium requests/month (Claude Sonnet 5 / GPT-5.6 / Gemini 3.7 Pro), unlimited slow requests, Cursor Tab autocomplete, background agents. Annual lock-in secures pricing for 12 months even if monthly price increases. Best for: daily coders using Cursor as primary IDE.
Aliyun Bailian qwen-3.8-turbo free 7 days for new accounts (unlimited)
New Aliyun Bailian accounts get qwen-3.8-turbo free for 7 days (unlimited tokens). After 7 days: ¥0.5/¥1.5 per 1M input/output. Compatible with OpenAI/Anthropic SDKs. Best for: small dev projects, hobby learning, quick API trials. Note: limited to qwen-3.8-turbo only, NOT qwen-3.8-max/qwen-3.8-coder.
GitHub Copilot Pro Annual — save $40 vs monthly ($100/yr vs $10/mo)
GitHub Copilot Pro annual subscription saves $20 vs paying monthly ($100/year vs $120/year). Same features: unlimited code completions, Next Edit Suggestions, unlimited chat with GPT-5.4/o3, 1500 AI Credits/month (1000 base + 500 flex), GitHub Copilot Workspace, Copilot Extensions. Refundable within 30 days.
Cursor Pro for Students — 50% off annual ($96/year)
Verified students (.edu email) get Cursor Pro annual at $96/year (50% off $192). Same features as Cursor Pro: 500 fast premium requests/month, 500 slow premium requests/month, unlimited Tab completions, unlimited Background Agents (beta), Bugbot (beta). Limited-time verification via SheerID.
GLM-5.3-Flash 5-fold promo ended 9/9 — reverts to ¥0.8/M input
GLM-5.3-Flash ran a limited 5-fold promo from 8/12-9/9 at ¥0.4/M input. As of 9/10 the price reverts to the standard ¥0.8/M input rate. Subscribe to TokenRhythm or watch this page for future flash promos.
DeepSeek V4.1 Flash valley pricing: ¥1/M input during off-peak hours
Starting 9/10, DeepSeek V4.1 Flash introduces time-of-use pricing. Off-peak (00:00-08:00 GMT+8 weekdays + all weekends) input is ¥1/M; peak (08:00-24:00 weekdays) input is ¥2/M. Output ¥4/M flat. First Chinese model with tiered pricing.
DeepSeek V4-Pro 75% Discount Made Permanent (Was Set to Expire 5/31)
DeepSeek made the 75% promotional discount permanent. V4-Pro now permanently at $0.435 / $0.87 per 1M tokens (cache miss input / output). Cached input: $0.003625 per 1M (90% off). 1M context window, up to 384K output per request. 8/16 added peak/off-peak billing: peak $1.32/$3.96, off-peak (17h) $0.66/$1.98.
Anthropic Cancels Sonnet 5 Sep 1 Price Hike: $2/$10 Permanent
Anthropic announced on Aug 10, 2026 that the planned Sep 1 price hike for Claude Sonnet 5 (from $2/$10 to $3/$15 per 1M tokens) is cancelled. The intro rate of $2/$10 is now the standard rate. Note: Sonnet 5 uses a new tokenizer that charges +30% tokens for identical text, so effective cost may still be higher than Sonnet 4.6.
GLM-5.3 API Launch - Same Price as GLM-5.2
Zhipu's GLM-5.3 launched on Aug 19, 2026 with API access. Pricing matches GLM-5.2 at $1.40/$4.40 per 1M tokens. AA Intelligence Index of 60 ties with Claude Fable 5 and GPT-5.6 Sol. Weights open-sourced Aug 28.
TokenRhythm 1M context API at ¥1/¥1 — DeepSeek V4 + GLM-5.2 + qwen3.7-max bundle
TokenRhythm aggregates 4 1M-context API models at unified ¥1/¥1 per 1M input/output: deepseek-v4-pro-0813 (1.6T A49B, 95.3% tokenizer ratio), glm-5.2 (Zhipu), qwen3.7-max (Alibaba). DeepSeek V4 cache read at ¥0.2/M. Compare to direct DeepSeek V4-Pro ¥4/¥12 (save 75%) and direct GLM-5.2 ¥0.8/¥2 (still cheaper direct). Best for: RAG workloads on budget. OpenAI/Anthropic SDK compatible.
Gemini 3.7 Flash 1M context now GA — $0.075/$0.30 flat (no long-context surcharge)
Google Gemini 3.7 Flash 1M context window now generally available. $0.075/M input, $0.30/M output flat — no 2x long-context surcharge above 128K (unlike GPT-5.6 / Claude Opus 4.8). Needle-in-haystack RAG benchmark: 94.3% retrieval accuracy at 1M. Best for: cost-sensitive RAG, large document analysis, foreign deployment. 4.3% retrieval accuracy > DeepSeek V4-Pro (93.1%).
GLM-5.2 1M context ¥0.8/¥2 — domestic cheapest 1M token API
Zhipu GLM-5.2 1M context window at ¥0.8/M input, ¥2/M output. Domestic (China-region) cheapest 1M context API. Long-context >128K: flat (no surcharge). Needle-in-haystack: 89.4% retrieval accuracy. Rival: DeepSeek V4-Pro ¥4/¥12 (5x more expensive), qwen-3.8-max ¥3/¥9 (3.75x). Best for: heavy Chinese RAG, large doc summarization, internal agent deployments.
GPT-5.6 Sol 1.05M context $2/$12 — long-context >272K adds 2x surcharge
OpenAI GPT-5.6 Sol tier: $2/M input, $12/M output at <272K input. Above 272K input: 2x long-context surcharge applies ($4/M input, $24/M output). Total 1.05M context window. Best for: high-quality RAG, agentic coding where context quality > cost. Compare: Gemini 3.7 Flash flat $0.075/M (40x cheaper for >272K).
Anthropic Claude Code Cloud Sessions GA — Pro $100 / Max $250 free credits (claim by 10/7)
Anthropic launched Claude Code cloud sessions on Sep 24: Pro subscribers get $100 in credit, Max subscribers get $250. Use --cloud flag to run parallel coding tasks on managed VMs even when your local workstation is offline. Claim credits by Oct 7; expire Nov 5. Access via claude.ai/code, mobile app, desktop, or terminal.
Aliyun Bailian decision-model-preview — 限时免费 API（高频业务判断）
Aliyun Bailian launched decision-model-preview on Sep 24: model ID for high-frequency business decision routing (risk control, ticket triage, marketing approval). Full API access free for preview period. Unified gateway with Qwen3.8 / DeepSeek V4 / GLM-5.2 / Kimi K2.7 on the same account.
豆包工作（字节）30 天免费订阅 — 新老用户均可领取，已付费用户自动延期
ByteDance's Doubao Work (豆包工作) launched Sep 23: 30-day free subscription for both new and existing users (paid users auto-extended). Workbench + AI assistant included. Released less than 2 months ago; AI office subsidy war has started (vs Tencent WorkBuddy / Feishu / DingTalk).
OpenRouter Free — 24 :free models + $10 lifetime top-up unlocks 5x rate limits
OpenRouter free tier: 24 :free models online (Gemini-2.0-Flash / Qwen3.8-27B / GLM-5.2 / NVIDIA Nemotron 3 Ultra 1M / Llama-4 etc.). Free base: 200 req/day. $10 lifetime top-up unlocks 5x rate limits. Context range 33K-1M. No credit card required.
Qwen3.8-Omni-Flash 9/26 GA — 1M context native multimodal (text+image+audio+video), audio cost down 98%
Alibaba Qwen team launched Qwen3.8-Omni-Flash on Sep 26: native multimodal at the model level (not stitched ASR + CV), 1M context window, audio input cost down 98%+, audio+video combo down 93%+. Beats Gemini 3.8 Flash on WildClawBench-MM 71.0 vs 58.9, DailyOmni 85.1 vs 84.0, SpotSoundBench 67.2 vs 39.7. Pricing: ¥1.5/¥4.7 per 1M (domestic) / $0.15/$0.47 per 1M (international). Supports 113 languages, stereo + 4-channel spatial audio.
Kimi K3 1M Context API — ¥3/¥15 per 1M, AAII 61 GPT-5.6 Sol 同级, 2.8T params MoE
Moonshot AI Kimi K3: 2.8T params MoE, 1M context window, AA Intelligence Index 61 (matches GPT-5.6 Sol). Pricing ¥3/¥15 per 1M (CN domestic) / $3.3/$16.5 (NEAR AI Cloud). K2.5/moonshot-v1 8/31 deprecated, migrate to K3. Best Chinese long-document understanding.
Claude Opus 5.5 (9/22) — 1M context, $3/$15 per 1M (40% cheaper than Opus 5)
Anthropic launched Claude Opus 5.5 on Sep 22: 1M context window, $3/$15 per 1M (40% cheaper than Opus 5's $5/$25). Matches Opus 5 quality at 60% price. CurveBench 3.2 parity with Opus 5. Cache write $3.75, read $0.30.
GPT-6 Sol/Luna 9/25 launched — prices halved, GPT-6 Astra 1.1M context preview
OpenAI launched GPT-6 Sol and Luna on Sep 25: prices slashed 50%. GPT-6 Astra 1M+ context preview available via waitlist for ChatGPT Pro. Sol: $2/$12 (was $4/$20, -50%). Luna: $0.10/$0.60 (was $0.20/$1.20, -50%). 1.05M-1.1M context. Astra features multi-Agent long-horizon collaboration.
Cursor Pro Annual Plan — $192/yr ($16/mo), save $48/yr vs monthly billing
Cursor Pro annual subscription officially priced at $192/year ($16/month equivalent), a 20% discount vs monthly $20. Includes AI code completion, multi-file editing, Cmd+K quick edit, AI Agents, 500 fast premium requests/month (GPT-5.6, Claude Opus 4.8, Gemini 3 Pro, Composer 1.6). Best for solo developers, indie hackers, and freelance engineers who need the latest AI coding tools at minimal cost.
ChatGPT Plus Annual Plan — $156/yr ($13/mo), save $24/yr vs monthly
OpenAI ChatGPT Plus annual plan $156/year ($13/month equivalent), saving $24/yr vs monthly $20. Includes GPT-5.6 main access, GPT-6 Sol/Luna preview (where eligible), DALL-E 3 image generation, Advanced Data Analysis, Code Interpreter, Browse with Bing retired (replaced by Web Search), GPTs custom creation, 50 messages/3hrs on GPT-5.6. Best universal AI assistant for solo users.
Anthropic Claude Max 5x Annual — $100/yr ($8.3/mo), 5x Pro usage, perfect for Claude Code power users
Anthropic launched Claude Max 5x on Sep 24 — annual subscription priced at $100/year ($8.3/month equivalent), delivering 5x the usage of Claude Pro ($20/month). Total annual value vs Pro ($240/yr): save $140/yr. Perfect for Claude Code power users, heavy AI Agent runners, and developers doing >100M tokens/month. Includes Claude Code Cloud Sessions Pro $100/$250 credits (claim by 10/7, expire 11/5) — let Claude AI run tasks on the cloud while local machine sleeps.
GitHub Copilot Pro Annual — $100/yr ($8.3/mo), save $40/yr vs monthly $10
GitHub Copilot Pro annual subscription $100/year ($8.3/month equivalent), saving $40/yr vs monthly $10. Includes AI code completion across VS Code / JetBrains / Visual Studio, Copilot Chat, Copilot Workspace Beta, 300 premium requests/month. Integrates with GPT-5.6, Claude Sonnet 5, Gemini 3 Pro models. Best for IDE-native developers who prefer Microsoft's stack.
China Domestic LLM Trio Stack — DeepSeek V4.1 Flash + GLM-5.3-Flash 5折 + Qwen3.8-Flash 1M = ¥450/month covers 100M tokens
Comprehensive Chinese domestic LLM combination stack at ¥450/month covers 100M tokens: (1) DeepSeek V4.1 Flash off-peak ¥0.25/M input + ¥1/M output for bulk batch; (2) Zhipu GLM-5.3-Flash 5折 discount ¥0.4/¥1.4 per 1M for general inference (extended through 10/7); (3) Alibaba Qwen3.8-Flash 1M ¥1.2/¥4 per 1M for long-context RAG. Total monthly ≈ ¥450 for 100M tokens (mix-30/40/30). All accessible through TokenRhythm aggregator for unified API + unified billing.
Qwen3.8-Omni-Flash 9/26 发布：原生多模态 1M 上下文 ¥1.5/¥4.7 per 1M（音频成本下降 98%）
Alibaba Qwen3.8-Omni-Flash launched 9/26: native multi-modal (text+image+audio+video) with 1M context. ¥1.5/¥4.7 per 1M (international $0.15/$0.47). Audio costs down 98%+. 113 languages. Replaces older Qwen3-Omni.
Kimi K3 2.8T 1M 上下文 9/25 上线：¥3/¥15 per 1M（首个 1M 上下文国产 MoE 模型）
Moonshot's first 1M-context model. ¥3/¥15 per 1M (~$0.43/$2.14). 2.8T total params, MoE architecture. Replaces 256K-only K2.7 lineage. Suitable for long-context research and code reasoning.
国产 Top 5 模型组合订阅 ¥450/月 100M token：DeepSeek V4.1 Flash 谷时 + GLM-5.3-Flash 5折 + Qwen3.8-Flash 1M + Kimi K3 1M + 豆包 2.0 Pro
Comprehensive domestic LLM stack for September 2026: combines the 5 cheapest/fastest domestic models into a single workload strategy. Covers everyday chat (DeepSeek V4.1 Flash valley ¥0.25/M), fast Q&A (GLM-5.3-Flash 5-fold ¥0.4/M until 10/7), 1M context batch (Qwen3.8-Flash 1M ¥1.2/M), long-context reasoning (Kimi K3 1M ¥3/M), and multimodal (Doubao 2.0 Pro ¥1/M). For 100M tokens/month batch run: ¥62.5 (DeepSeek valley) + ¥90 (GLM 5-fold) + ¥260 (Qwen 1M) + ¥900 (Kimi 1M) + ¥350 (Doubao) ≈ ¥1500-2000, plus subscription ¥450/mo flat = unlimited generation. Best stack for indie developers and SMBs.
Qwen3.8-Omni-Flash ¥1.5/¥4.7 per 1M（9/26 新发布）= 国内首个原生多模态 1M 上下文，音频成本 -98%
Alibaba Tongyi Qwen3.8-Omni-Flash launched 9/26 at ¥1.5 input / ¥4.7 output per 1M tokens — first domestic native multi-modal 1M-context model. Native multi-modal (text+image+audio+video), not ASR+CV stitched. WildClawBench-MM 71.0 (vs Gemini 3.8 Flash 58.9). Audio input cost -98%, audio+video combination -93% vs prior Qwen3.8-Flash. Supports 113 languages, stereo + 4-channel spatial audio. Compared to Gemini 3.8 Flash $0.075/$0.30 input: domestic pricing 1x cheaper input, 1.5x more expensive output; compared to GLM-5.3-Flash 5-fold ¥0.4/¥1.4 (text-only): significantly more capable for audio workloads.
DeepSeek V4-Pro-0813 1M ¥1/¥2 per 1M（8/12 GA）= 国产按量输入最便宜，MoE 1.6T A49B
DeepSeek V4-Pro-0813 GA 8/12 at ¥1 input / ¥2 output per 1M tokens, 1M context, 1.6T total params / 49B activated MoE. The cheapest domestic 1M-context model by input price; output ¥2/M vs Kimi K3 1M ¥15/M (7.5x more expensive) vs Qwen3.8-Flash 1M ¥4/M (2x). For long-context RAG workloads: 1M 大海捞针 93.1% accuracy, second only to Gemini 3.7 Flash 1M 94.3%. For 100M tokens/month batch run: ¥1600 (vs Kimi K3 ¥9000 if output-heavy). Best cheap 1M domestic pick.
GLM-5.2 ¥1/¥2 per 1M（200K 上下文国产最便宜 1M）= 355B A32B MoE，国产 1M 上下文 GA 路径
Zhipu GLM-5.2 at ¥1 input / ¥2 output per 1M tokens, 200K context, 355B total / 32B activated MoE. The cheapest 200K-context domestic API by both input and output (tied with DeepSeek V4 Flash ¥1/¥2 peak). For RAG workloads under 200K: 89.4% accuracy in 1M 大海捞针 style benchmarks. 100M tokens/month batch: ¥280 — between Qwen3.8-Flash 1M ¥260 and TokenRhythm ¥200. Best mid-context domestic pick (200K is enough for 80% of RAG use cases).
Kimi K3 2.8T 1M ¥3/¥15 per 1M（9/25 上线）= 月之暗面首个 1M MoE，AAII 87.99 持平 GPT-5.6 Sol
Moonshot AI Kimi K3 launched 9/25 with 2.8T total params / 32B activated MoE, 1M context (first Kimi 1M model, replacing 256K-only K2.7). ¥3 input / ¥15 output per 1M. AAII 87.99 = GPT-5.6 Sol level. Long-context RAG 1M 大海捞针 89.7%. Most expensive output among domestic 1M models (¥15 vs ¥2-4 for DeepSeek/GLM/Qwen), but premium reasoning + extended thinking make it best for research/agent workloads. Kimi K2.5/moonshot-v1 deprecated 8/31 — must migrate to K3.
阿里百炼 Token Plan 三档 ¥9.9/¥39/¥99 月订阅 = 国产 OpenAI 月订阅最低门槛
Aliyun Bailian (Alibaba Cloud Model Studio) Qwen Token Plan 3 tiers: (1) Qwen Lite ¥9.9/mo — 10M tokens/mo Qwen3.8-Turbo/Qwen3.8-Flash access; (2) Qwen Standard ¥39/mo — 50M tokens/mo + Qwen3.8-Max + Qwen3.8-Omni; (3) Qwen Pro ¥99/mo — 200M tokens/mo + Qwen3.8-Max unlimited reasoning + Qwen3.8-Omni. Best cheap entry for Qwen family. For domestic LLM stackers: ¥99 Pro + DeepSeek V4.1 Flash + GLM-5.3-Flash 5-fold covers 95% workloads.
Gemini 4 Argon 1M Output + DeepSWE 77.9% SOTA — Fairwind 限量访问
Google 9/30 发布 Gemini 4 Argon：单次输出 1M tokens（Gemini 3 仅 64k），DeepSWE v1.1 = 77.9% 创 SWE-bench Verified 新 SOTA（Opus 5.5 74.2% / GPT-6 Astra 74.1%），CWE-bench v1 与 Gemini 3.8 Flash Cyber 并列第一 68%。介绍期 API $2/$10 per 1M，首批仅向 Fairwind 项目的可信网络防御者开放。
白宫超级智能协议 — 10 家前沿 AI 厂商 CEO 联合签署
特朗普 9/29 签署行政令将联邦文件中 "AI" 改为 "Super Intelligence"；同日与 OpenAI/Anthropic/Google/Meta/Nvidia/xAI/SpaceX/Palantir/Microsoft/Amazon 等 10 家 CEO 签署 "白宫超级智能协议：前沿责任联合承诺"（morally binding）。OpenAI 同步回溯 2026-01 以来智能体活动日志。
TokenRhythm 1M ¥500 Credit Extended to 10/31 — Cheapest 1M Aggregator
TokenRhythm 1M aggregator extends ¥500 trial credit to 10/31/2026 (originally 10/4, extended 9/30). Pricing remains ¥1 input / ¥1 output per M tokens with 1M context unified access to 11+ vendors (OpenAI/Anthropic/Google/DeepSeek/GLM/Qwen/Kimi/MiniMax/Mimo/xAI/TokenRhythm native). 100M token batch run = ¥200 + ¥500 credit = covers 500M tokens first month free. vs Gemini 3.7 Flash 1M (¥2.2/¥8.7) is 55% cheaper input, vs Claude Opus 4.8 1M (¥110/¥550) is 99% cheaper. Best entry point for cross-vendor arbitrage and unified billing.
Aliyun Bailian Qwen Standard — ¥39 First Month ¥27.3 (30% off, Oct only)
Aliyun Bailian Qwen Standard monthly subscription: ¥39/mo standard price, but first-month 30% off for new subscribers in Oct 2026 = ¥27.3/mo. Includes 50M tokens + Qwen3.8-Max access. After first month, regular ¥39/mo. Best value for Chinese developers wanting Qwen3.8-Max 1M context + Lite + Standard tier without annual commitment. ¥11.7 saved first month covers full Standard tier pricing.
OpenRouter — $10 Lifetime Credit + 24 :free Models
OpenRouter: register and get $10 lifetime credit (unlocks 5x rate limit) + access to 24 :free models including Gemini-2.0-Flash, Qwen3.8-27B, GLM-5.2, Nemotron 3 Ultra 1M, Llama-4. Best entry for developers wanting to test multiple vendors without committing to each. $10 = ~50M tokens of mid-tier API. Permanent credit never expires, free tier stays accessible even after credit runs out.
DeepSeek V4-Pro 1M Context — ¥2/¥8 per M (Cheapest 1M API)
DeepSeek V4-Pro-0813-31 with full 1M context window: ¥2 input / ¥8 output per 1M tokens (best price among 1M models). needle-in-haystack 99.5% accuracy (10/10 benchmark), 1.2s first-token latency for 500K input, 100% 30-round continuous conversation recall. Combined with V4.1 Flash valley pricing (¥0.25/M input 0-6/16-24), offers the cheapest 1M long-context API for Agent / code repo / document analysis workloads. Stacks with Oct recharge ¥500 get ¥600 promo.
GLM-5.3-Flash 1M — ¥0.8/¥2.8 per M (Recovered from 5折)
Zhipu GLM-5.3-Flash with 1M context window: ¥0.8 input / ¥2.8 output per 1M tokens (standard price after 10/7 5折 promo ended). needle-in-haystack 99.1%, 1.8s first-token for 500K input, tool-calling 99.4% success rate. Best value pick for Chinese developers wanting domestic 1M context + low latency + tool-use reliability. Pairs with GLM Coding Plan Max (¥120/mo for 2000 GLM-5.3-Flash + 50 Sonnet 5 calls).
Kimi K3 2.8T 1M Monthly — ¥299/mo First Month ¥239 (8折)
Moonshot AI Kimi K3 (2.8T parameters) with 1M context window: ¥299/mo monthly subscription with 8折 first-month promo = ¥239/mo. needle-in-haystack 99.7% SOTA (best recall among 1M models), 2.4s first-token for 500K input, ¥3/¥9 per 1M token overage. Equivalent to 99M tokens/month full-context use. Best for K2.7 upgrade users needing best-in-class long-context recall.
Claude Opus 4.8 1M — ¥110/¥550 per M (Premium Long-Context Coherence)
Anthropic Claude Opus 4.8 with full 1M context window: ¥110 input / ¥550 output per 1M tokens (premium tier, ~55x DeepSeek V4-Pro). needle-in-haystack 99.6%, 2.8s first-token for 500K input, BEST long-context coherence for writing/code/refactoring tasks. Pricing tiers: 0-200K standard (¥18/¥90), 200K-1M 9x multiplier. Use Max 5x ($100/yr) for fixed-cost Opus 4.8 access. Best for premium writing / complex code agent / extended research where coherence > cost.
TokenRhythm 1M — ¥1/¥1 per M Cross-Vendor Aggregator (¥500 Credit to 10/31)
TokenRhythm 1M aggregator: ¥1 input / ¥1 output per M tokens, 1M context unified access to 11+ vendors including all 5 1M-context models (DeepSeek V4-Pro / GLM-5.3 / Kimi K3 / Gemini 4 Argon / Claude Opus 4.8). ¥500 trial credit extended to 10/31. Switch between 1M models on-the-fly without re-integrating APIs. 100M token batch = ¥200 + ¥500 credit = 500M tokens free first month. BEST entry for cross-vendor 1M model arbitrage.
Cursor Pro Annual — $192/yr ($16/mo, save $48 vs monthly)
Cursor Pro annual subscription at $192/year ($16/mo equivalent) saves $48/year vs monthly $20/mo. Includes 500 fast premium requests/month with GPT-5.6 Luna + Claude Sonnet 5 + Gemini 3.7 Flash auto-switching. Best IDE subscription baseline for 2026 Q4.
Q4 2026 Best AI Subscription Stack — Cursor Pro + GLM Coding Max + TokenRhythm 1M < $50/mo
Q4 2026 best subscription combo covering 99% of solo dev workflows: Cursor Pro Annual $16/mo + GLM Coding Max ¥120/mo + TokenRhythm 1M ¥200/100M. Total < $50/mo comprehensive coverage including IDE + 2000 GLM-5.3 calls + 50 Sonnet 5 + 100M 1M-context aggregator tokens.
DeepSeek V4-Pro — ¥2/¥8 per 1M tokens (Cheapest 1M Pro, 30-turn 100% recall)
DeepSeek V4-Pro 1M context API at ¥2 input / ¥8 output per 1M tokens. 500K-1M continuous 30-turn conversation 100% recall; first-token latency 1.2-2.8s (fastest in class). W42 cheapest 1M Pro pick.
DeepSeek V4.1 Flash Valley — ¥0.25/¥1 per 1M (0-6/16-24, W42 cheapest batch)
DeepSeek V4.1 Flash off-peak pricing at ¥0.25 input / ¥1 output per 1M tokens during 0-6 / 16-24 hours. Perfect for batch workloads (data cleaning / log analysis / bulk translation). W42 cheapest off-peak API pick.
GLM-5.3-Flash — ¥0.8/¥2.8 per 1M (1M context, 99.4% tool-use, W42 domestic flash #1)
Zhipu GLM-5.3-Flash 1M context API at ¥0.8 input / ¥2.8 output per 1M tokens (post double-festival 10/7). 99.4% tool-use success rate. W42 cheapest domestic flash pick.
Qwen3.8-Flash 1M — ¥1.2/¥4 per 1M (2nd cheapest domestic 1M Flash, WildClawBench 61.9)
Aliyun Qwen3.8-Flash 1M context API at ¥1.2 input / ¥4 output per 1M tokens. Second cheapest domestic 1M Flash pick. WildClawBench 61.9; strong for Chinese long-form RAG.
TokenRhythm 1M Aggregator — ¥1/¥1 per M, ¥500 credit till 10/31
TokenRhythm 1M aggregator at ¥1 input/¥1 output per 1M tokens, unified 1M context entry across 5 vendors. ¥500 signup credit extended to 10/31. Best cross-vendor arbitrage pick.
DeepSeek V4-Pro-0813 1M Domestic-Pro #1 (¥2/¥8) — Cheapest 1M Pro in W42
DeepSeek V4-Pro-0813 (1.6T A49B MoE, 1M context, 384K max output) ¥2/¥8 per 1M tokens. 30-turn 100% recall, 1.2-2.8s first-token (fastest). Swe-bench Verified 79.4% (domestic #1). OpenAI/Anthropic API compatible via TokenRhythm 1M aggregation.
Zhipu GLM-5.3-Flash 1M (¥0.8/¥2.8) — Domestic Flash #1, 99.4% Tool-Use
Zhipu GLM-5.3-Flash 1M context (post 10/7 returned to list price) ¥0.8/¥2.8 per 1M tokens. Domestic Flash #1. 99.4% tool-use success (BFCL-v3, complex function nesting — domestic #1). Combine GLM Coding Max ¥120/mo for 2000 calls + 50 Sonnet 5.
Moonshot Kimi K3 2.8T (¥3/¥15) — 99.7% RAG Recall Domestic SOTA
Kimi K3 2.8T A48B MoE (9/25 GA, first domestic 1M context MoE) ¥3/¥15 per 1M tokens. 99.7% needle-in-haystack recall (500K doc, 50 random queries — domestic SOTA, beats Claude Opus 4.8 99.6%). Tool-use 96.8%. SWE-bench 71.8%.
Alibaba Qwen3.8-Flash 1M (¥1.2/¥4) — Domestic 1M Flash #2, WildClawBench-MM 61.9
Aliyun Qwen3.8-Flash 1M context (2.4T A95B MoE) ¥1.2/¥4 per 1M tokens. Domestic 1M Flash #2 (ranked just below GLM-5.3-Flash ¥0.8). WildClawBench-MM 61.9. Token-plan subscription: Qwen Lite ¥9.9 / Standard ¥39 (7折 ¥27.3) / Pro ¥99 (5折).
Zhipu GLM Coding Max ¥120/mo — 2000 GLM-5.3-Flash + 50 Sonnet 5 Calls/Month
Zhipu GLM Coding Max IDE subscription ¥120/month: 2000 GLM-5.3-Flash calls + 50 Anthropic Sonnet 5 calls per month. Domestic IDE Coding Plan best-value. Tool-use 99.4% (BFCL-v3). Combine with TokenRhythm 1M aggregator for 5× model switch.
TokenRhythm 1M ¥1/¥1 + ¥500 Credit — Cross-Vendor 5-Model 1M Aggregator
TokenRhythm 1M aggregator: DeepSeek V4-Pro / GLM-5.3-Flash / Qwen3.8-Flash / Kimi K3 / Doubao 2.0 Pro all at ¥1/¥1 per M tokens. ¥500 sign-up credit extended to 10/31. OpenAI/Anthropic-compatible API. 100M tokens batch ¥200.
DeepSeek V4-Pro 1M ¥2/M Input — 30-Turn 100% Recall Multi-Turn Agent SOTA
DeepSeek V4-Pro 1.6T MoE A49B reaches 100% needle-in-haystack recall across 30-turn agent benchmarks at 1M context. ¥2/M input + ¥8/M output. Production-tested at Replit (since 10/16) and Vercel AI Gateway (default route for TypeSafe Jev). 1.2s first-token at 500K input — fastest among 1M models. Best cost-vs-recall for long-horizon coding agents.
GLM-5.3-Flash 1M ¥0.8/M Input — 99.4% Tool Calling, Cheapest 1M + Tool-Use
GLM-5.3-Flash A26B MoE achieves 99.4% tool-calling accuracy at 1M context (BFCL v3 leaderboard). ¥0.8/M input + ¥2.8/M output. 137x cheaper than Claude Opus 4.8 at ¥110/M. Best 1M model for tool-use agents on a budget. Combined with GLM Coding Max ¥120/mo (2000 calls/month) = $1 per call effective cost.
Kimi K3 2.8T 1M ¥3/M Input — 99.7% RAG SOTA, Largest Active MoE 1M Model
Kimi K3 2.8T total params / 48B active A48B achieves 99.7% needle-in-haystack recall at 1M context (top score). ¥3/M input + ¥15/M output. Best-in-class for Chinese long-document RAG (legal / academic / financial reports). Note: China MoCT 10/16 export controls may affect overseas self-hosting; API access unaffected.
Claude Opus 4.8 1M ¥110/M Input — 9.6/10 Coherence Premium Long-Context
Claude Opus 4.8 1M scores 9.6/10 coherence (LLM-as-judge) at 1M context — highest among 5-strong. ¥110/M input + ¥550/M output. Premium tier for: legal contracts with cross-clause references, full-codebase reasoning, long-form creative writing. Not cost-effective for batch; use Opus 4.8 only when coherence > cost.
TokenRhythm 1M Aggregator ¥1/M Flat — All 5 Strong 1M Models in One API, ¥500 Credit Until 10/31
TokenRhythm exposes DeepSeek V4-Pro + GLM-5.3-Flash + Kimi K3 + Gemini 4 Argon + Claude Opus 4.8 via one API at ¥1/M input flat. No model selection overhead; smart router picks best for prompt. ¥500 credit extended to 10/31. Default for new projects; pin specific vendor only for coherence-critical workloads.
Gemini 4 Argon 1M Output $2/$10 — DeepSWE 77.9% SOTA Single-Shot
Gemini 4 Argon produces 1M tokens in single output (most competitors cap at 200K). DeepSWE v1.1 = 77.9% SOTA on SWE-bench Verified (vs Opus 5.5 74.2% / GPT-6 Astra 74.1%). $2/M input + $10/M output. Native multimodal: image + text in unified 1M context. Competes with Elorian Glimpse (10/16 launch).
Meta Muse Spark Open 816B MoE — Free for non-commercial, multimodal 1M context, MLX accelerated
Meta released Muse Spark Open 10/16: 816B MoE multimodal model with 1M context window, MLX acceleration, Apache 2.0 license. Free for non-commercial use; commercial license available. Self-hostable on Apple Silicon (M3 Ultra+) for image+text workloads. Competitor to Gemini 4 Argon multimodal but free.
DeepSeek V4.1 Flash 1017 Build — MIT open-source, Terminal-Bench 4.0 = 42.3, 谷时 ¥0.25/M input
DeepSeek V4.1 Flash 1017 build released 10/17, Terminal-Bench 4.0 jumped to 42.3, weights MIT open-sourced. ¥0.25/M input at 谷时 (0-6/16-24 UTC) for bulk processing. The cheapest 100M token / month option now < ¥62.5. Critical for heavy agent stacks.
Google Antigravity CLI v1.4.1 — Free CLI with plugin verification and diff preview, integrates Gemini 4 Argon
Google released Antigravity CLI v1.4.1 10/17: command-line agent harness with plugin verification, diff preview, and Gemini 4 Argon integration. Free for all developers. Alternative to Claude Code CLI; works on macOS/Linux/Windows. Best for: terminal-first agent workflows.
GLM-5.3-Flash Standard ¥27.3/月 (首月 7 折) — 100 次 GLM-5.3-Flash 调用, 200K context, tool-use
智谱 GLM-5.3-Flash Standard launched 10/15: ¥27.3/月 (7 折首月 = ¥19.1). Includes 100 GLM-5.3-Flash calls (200K context, 99.4% tool-use success). Lighter entry than GLM Coding Max ¥120/月 (2000 calls). Best for: solo devs starting out.
Cursor Pro Annual + Cloud Agent Canvases Beta (10/16) — $16/mo with remote agent execution
Cursor added Cloud Agent Canvases + Environments API Beta on 10/16: run agents on remote cloud infrastructure directly from the IDE. Annual $192/yr = $16/mo (save $48/yr). Combines IDE + cloud agent execution for heavy workflows.
DeepSeek V4.1 Flash 1017 Build ¥0.25/¥1 per 1M (谷时 0-6/16-24，MIT 开源，Terminal-Bench 4.0=42.3)
DeepSeek V4.1 Flash 1017 build released 10/17, weights MIT open-sourced, Terminal-Bench 4.0 jumped to 42.3. ¥0.25/M input at 谷时 (0-6/16-24 UTC) for bulk processing — the cheapest 100M token / month option now < ¥62.5. Critical for heavy agent stacks.
GLM-5.3-Flash ¥0.8/¥2.8 per 1M (1M 上下文，99.4% 工具调用，W43 国产 Flash #1)
GLM-5.3-Flash 1M context: ¥0.8 input / ¥2.8 output per 1M tokens. 99.4% tool-calling success rate. W43 国产 1M Flash 第一名（vs Qwen3.8-Flash ¥1.2, Kimi K3 ¥3）。搭配 GLM Coding Max ¥120/月（2000 次 + 50 Sonnet 5）或新 GLM-5.3-Flash Standard ¥27.3/月（100 次）。
Qwen3.8-Flash 1M ¥1.2/¥4 per 1M (国产 1M Flash 第二便宜，WildClawBench-MM 61.9)
Qwen3.8-Flash 1M context: ¥1.2 input / ¥4 output per 1M tokens. W43 国产 1M Flash 第二便宜（vs GLM-5.3-Flash ¥0.8）。WildClawBench-MM 61.9。阿里百炼 Qwen Standard ¥27.3 首月 7 折（10/15）。搭配 Qwen Coding 套餐 ¥80/月（800 次）。
TokenRhythm 1M 聚合 ¥1/¥1 per M, ¥500 体验金延期至 10/31 (5 款 1M 模型统一入口)
TokenRhythm 1M aggregator: ¥1 input + ¥1 output per 1M tokens (flat). 5-strong unified API (DeepSeek V4-Pro / GLM-5.3-Flash / Kimi K3 / Gemini 4 Argon / Claude Opus 4.8). ¥500 signup credit extended to 10/31 — **W43 最后 12 天**，错过等 11 月。
Kimi K3 1M ¥3/¥15 per 1M (99.7% 大海捞针召回 SOTA，国产首个 1M MoE A48B)
Kimi K3 1M context: ¥3 input / ¥15 output per 1M tokens. 99.7% needle-in-haystack recall (W43 RAG SOTA). 国产首个 1M MoE A48B（2.8T 总参）。100M token/月 ¥900 比 V4-Pro ¥1600 略便宜但 RAG 准确率第一。**W43 新风险**：商务部 10/16 出口管制清单在列。
DeepSeek V4-Pro 1M ¥2/¥8 per 1M tokens — W43 国产 1M Pro 最便宜，连续 2 周榜首 + SWE-bench 国产 #1
W43 (10/20-10/26) cheapest 1M Pro stays ¥2/¥8 per 1M; 30-round 100% recall on 500K-1M context; SWE-bench Verified 79.4% W43 domestic #1. CursorBench 综合 4 子榜领先. First-token latency 1.2-2.8s (W43 fastest).
DeepSeek V4.1 Flash 1017 MIT Open Source + Valley ¥0.25/¥1 per 1M (0-6/16-24 UTC) — 全榜最便宜跑批
W43 (10/20-10/26) new 1017 build on HuggingFace under MIT license; open weight + commercial use allowed. Valley pricing ¥0.25/M input ¥1/M output (0-6 / 16-24 UTC). Peak ¥1/M ¥2/M. Cheapest 100M-token batch in market vs Kimi K3 ¥900 / Claude Opus ¥77000.
GLM-5.3-Flash 1M ¥0.8/¥2.8 per 1M tokens — W43 国产 Flash #1 + 99.4% 工具调用 + 5 折 (10/14-10/28)
W43 (10/20-10/26) Zhipu's 5-折 promo extended 10/14-10/28 brings input from ¥0.8 to ¥0.4 (new users); 1M context + 99.4% BFCL-v3 tool-use (domestic #1). Standalone GLM-5.3-Flash Standard subscription ¥27.3/month (100 light calls/mo) also live.
GLM Coding Max ¥120/月 IDE Coding Plan #1 — 2000 次 GLM-5.3-Flash + 50 次 Sonnet 5 (国产 IDE Coding 性价比首选)
W43 (10/20-10/26) Zhipu's flagship IDE coding plan: ¥120/month includes 2000 GLM-5.3-Flash calls (most use cases) + 50 Sonnet 5 fallback for hard tasks. Domestic IDE Coding Plan #1 vs Cursor Pro $20/mo + Copilot Pro $10/mo.
Kimi K3 2.8T 1M ¥3/¥15 per 1M tokens — W43 1M RAG SOTA 99.7% + 8 折 (10/22-11/5 限时)
W43 (10/20-10/26) Moonshot launches Kimi K3 autumn promo: ¥3/M ¥15/M (1M context) + 8-折 for first-month ¥239/mo K3 Pro subscribers. 1M needle-in-haystack 99.7% (domestic SOTA) + 2.8T A48B MoE + SWE-bench 71.8%.
TokenRhythm 1M ¥500 体验金延期至 10/31 + 国产 5 强 ¥1/¥1 跨厂商套利 (W43 最后 11 天)
W43 (10/20-10/26) TokenRhythm extends ¥500 trial credit to 10/31 (last 11 days). 5-Model unified entry at ¥1/M input ¥1/M output. Includes DeepSeek V4-Pro / GLM-5.3-Flash / Qwen3.8-Flash / Kimi K3 / Doubao 2.0 Pro.
Cursor 10/21 Lineup Rebuild: Hobby/Pro Retired, Plus $8 / Premium $30 (4x) / Super $100 (15x) / Ultra $200 (50x)
Cursor restructured its entire personal lineup on 10/21. Hobby Free and Pro $20 are retired; the new ladder is Free $0 / Plus $8 (entry paid) / Premium $30 (4x Plus, marked most popular) / Super $100 (15x Plus, peak priority) / Ultra $200 (50x Plus, earliest access to all models). Teams is $30+/user/month (4-50x Plus) with centralized billing, usage analytics and SSO; invoicing/PO, SCIM and audit logs are Enterprise-only via contact sales.
Claude Sonnet 5 Stays $2/$10 — Anthropic Cancels the Planned September Hike to $3/$15
Anthropic's official pricing page confirms Claude Sonnet 5 remains at $2 input / $10 output per 1M tokens; the previously announced September increase to $3/$15 has been cancelled. Full ladder verified 10/21: Opus 5.5 $4/$20, Opus 5 / 4.8 / 4.7 / 4.6 / 4.5 $5/$25, Sonnet 5.5 $2/$10, Sonnet 4.6 / 4.5 $3/$15, Haiku 5.5 $0.10/$0.50 (<=100K), Haiku 4.5 $1/$5, Fable 5.1 / Mythos 5.1 $10/$50. Note Claude 4.7+ uses a new tokenizer producing ~30% more tokens for the same text, so cross-vendor comparisons must be normalized.
DeepSeek V4-Pro Valley/Peak Split Confirmed: $0.66/$1.32 input, $1.98/$3.96 output (off-peak/peak)
Official pricing verified 10/21: deepseek-v4-pro (V4-Pro-0813) cache hit $0.022/$0.044, cache miss $0.66/$1.32, output $1.98/$3.96 per 1M tokens (off-peak / peak). deepseek-flash (= V4.1-Flash-1017) cache hit $0.003/$0.006, cache miss $0.15/$0.30, output $0.60/$1.20. Peak hours are only Mon-Fri 01:00-04:00 and 06:00-10:00 UTC (excluding Chinese public holidays); all other hours including weekends and holidays are off-peak at roughly half the peak price.
GLM Coding ¥80/¥120 vs Token Plan ¥39/¥99 — Which Subscription Shape Fits Your Workload in W43?
Coding Plan (per-request) vs Token Plan (per-token bucket) vs API aggregator — W43 comparison. GLM Coding Pro/Max at ¥80/¥120 per month bundles 2000 GLM-5.3-Flash calls plus 50 Sonnet 5 calls, and pairs with GLM-5.3-Flash's 99.4% BFCL-v3 tool-use success rate. Alibaba Bailian token_plan: Lite ¥9.9 first month, Standard ¥39 (7折 = ¥27.3, unlimited Qwen3.8-Flash 1M), Pro ¥99 (5折 = ¥49.5, all models + 5000 calls). TokenRhythm aggregates at ¥1/¥1 per M. Rule of thumb: <500 IDE requests/day -> Coding Plan; stable single-vendor long context -> Token Plan; cross-vendor batch/agent work -> aggregator.
GitHub Copilot Credit Pricing Verified 10/21: Pro $10 (+$15 credits +$5 flex) / Pro+ $39 / Max $100 — Business & Enterprise Now Contact-Sales Only
Official plans page verified 10/21: Free $0, Pro $10/month including $15 GitHub AI Credits plus $5 flex, Pro+ $39 including $70 plus $31 flex, Max $100 including $200 plus $100 flex. 1 AI Credit = $0.01. Inline code completion and Next Edit suggestions are unlimited on all paid tiers. Business and Enterprise no longer publish list prices and require contact with sales. Usage alerts fire at 75%, 90% and 100% of the budget.
MiniMax M3 Priority Tier Confirmed at 1.5x Standard: ¥3.15/¥12.6 (<=512K) and ¥6.3/¥25.2 (>512K)
Official pricing verified 10/21: M3 standard tier is ¥2.1 input / ¥8.4 output / ¥0.42 cache for <=512K input tokens, and ¥4.2 / ¥16.8 / ¥0.84 for >512K. The priority tier (service_tier=priority) is exactly 1.5x standard: ¥3.15/¥12.6 and ¥6.3/¥25.2. M3.1-Flash is ¥1 input / ¥4 output / ¥0.1 cache. Note the official pricing page no longer lists Starter / Plus / Max / Ultra coding subscription tiers — only the M Plan quota deduction mechanism is described. Crossing the 512K input threshold doubles cost, so long-context workloads should split or use the aggregator.
DeepSeek Search/Fetch API stays free until Oct 31 — $0.75/1000 calls billing starts Nov 1
DeepSeek announced that per-request billing for its Search and Fetch APIs begins at 00:00 Beijing time on November 1, 2026, at $0.75 per 1,000 successful requests. Until then, server-side tool calls routed through DeepSeek Harness, Claude Code, or Codex remain free. Budget for a step-change in cost on agent pipelines that rely on server-side search.
Claude Sonnet 5 price hike officially cancelled — $2/$10 becomes the permanent standard price
Anthropic confirmed the previously announced September 1, 2026 increase from the $2/$10 introductory rate to $3/$15 will not take place. $2 per million input and $10 per million output tokens is now the standing price for both Sonnet 5 and Sonnet 5.5. The former August 31 promotional deadline is also void.
Claude Haiku 5.5 launches at $0.10/$0.50 per 1M for prompts under 100K — 10x cheaper input than Haiku 4.5
Claude Haiku 5.5 is priced at $0.10 input and $0.50 output per million tokens for requests of 100K tokens or less, making it the cheapest tier in the Claude lineup. Prompts above 100K step up to $0.50/$2.50. Cache hits are $0.01 per million on the short tier.
MiniMax-M3.1-Flash launches at ¥1/¥4 per 1M — new Chinese Flash-tier floor price
MiniMax added the M3.1-Flash tier at ¥1.00 input and ¥4.00 output per million tokens with ¥0.10 cache reads. For comparison, M3 at 512K or below is permanently half off at ¥2.10/¥8.40, so M3.1-Flash cuts the input price by another 52%. Priority service is 1.5x.
GLM-5.3-Flash limited-time 50% off at ¥0.40/¥1.40 per 1M — Opus 4.8-level capability at 1/40 the price
Zhipu's GLM-5.3-Flash is a natively multimodal model (image, video, file, text) whose standard price is ¥0.80/¥2.80 per million tokens. A limited-time 50% promotion brings it to ¥0.40/¥1.40 with cache hits at ¥0.115. Zhipai states some capabilities match Claude Opus 4.8. Batch API is 5x cheaper.
GPT-5.6 Sol $4/$20 promotional pricing locked through at least Nov 21, 2026 — Batch halves it to $2/$10
OpenAI's pricing page states GPT-5.6 Sol's $4 input / $20 output per million rate is promotional and available at least through November 21, 2026. Batch and Flex are both 50% of standard ($2/$10), while Fast mode is 2x. Requests over 272K input tokens double to $8/$30.