Q: W42 本周免费/优惠 API 汇总（10/13-10/19）有哪些值得申请的硬核福利？

W42 free/promo roundup (10/13-10/19). OpenAI GPT-6 Intelligent UI to free ChatGPT (GPT-5.5 retired 10/14); Apple Intelligence China free with iOS 27.3 in December; Anthropic Claude Marketplace 19 partners + METR independent evaluator access; xAI Grok Voice 3.0 $0.08/min trial; Alibaba Qwen Go personal agent free + Qwen Standard first month ¥27.3; OpenRouter 24 :free models (added DeepSeek V4-Pro 1M / GLM-5.3-Flash 1M / Kimi K3 on 10/16) + $10 lifetime credit; 5 open-source LLM launches on 10/15 (Meta Muse Spark Open 816B MoE / NVIDIA Nemotron 3.5 Super VL / Black Forest Labs FLUX.3 Image Klein 6B / IBM Granite Switch 4.1 / Mistral Timestral 209M); OpenAI Rosalind academic API credits; TokenRhythm 1M ¥500 credit extension to 10/31. Stackable hard-core <$30 starting pack: TokenRhythm ¥200 + OpenRouter $10 + Qwen Go free + Doubao work 30-day free + Hugging Face Pro $9 = $200 equivalent value.

Question 1

Which AI APIs are free or discounted this week (Oct 20-23, 2026)?

Accepted Answer

Five things qualify. (1) OpenAI's Decisions API went generally available on October 22 at $0.10 per million input tokens and $0.01 per million cached input tokens, and output tokens are not billed separately; a request accepts up to 400 questions and runs up to 10x faster than GPT-6 Luna through the Responses API. (2) iFlytek released Spark X2.6 on October 22 at CNY 0.9 input / CNY 0.08 cached input / CNY 3.6 output per million tokens with a 1M context window and a simultaneously open-sourced 4B distilled version. (3) xAI's grok-4.3 is $1.25/$2.50 with a 1M context window, 37% cheaper than Grok 4.7 which caps at 500K. (4) Anthropic documents free tool tiers: Web Fetch is free, Code Execution includes 1,550 free container hours per organization per month, and Web Search costs $10 per 1,000 successful searches. (5) Zhipu's GLM-5.3-Flash limited-time 50% discount at CNY 0.40/1.40 expires October 28, while MiniMax has made M3's 50% discount below 512K permanent at CNY 2.10/8.40 alongside the new M3.1-Flash at CNY 1.00/4.00.

Question 2

What are the biggest AI vendor price changes this week (Oct 15-22, 2026)?

Accepted Answer

Five moves matter. (1) DeepSeek Search and Fetch APIs start per-request billing at $0.75 per 1,000 successful calls on November 1, 2026; server-side tool calls through DeepSeek Harness, Claude Code, and Codex were free until then. (2) Anthropic cancelled the Claude Sonnet 5 price increase that was scheduled for September 1, 2026, making $2/$10 per million the permanent standard price for Sonnet 5 and Sonnet 5.5. (3) Claude Haiku 5.5 launched at $0.10/$0.50 per million for prompts up to 100K tokens, the cheapest tier in the Claude lineup and a 90% input-price cut versus Haiku 4.5. (4) OpenAI locked GPT-5.6 Sol at $4/$20 as a promotional rate available at least through November 21, 2026; Batch and Flex are 50% of that and requests over 272K input tokens double to $8/$30. (5) MiniMax launched M3.1-Flash at ¥1/¥4 per million while permanently extending M3's 50% discount at ¥2.10/¥8.40 for inputs up to 512K. Two smaller notes: Zhipu's GLM-5.3-Flash is at a limited-time 50% off to ¥0.40/¥1.40, and xAI's Grok 4.7 launched at the same $2/$6 as 4.6 but raised cache reads from $0.30 to $0.50.

Question 3

W43 Wednesday (2026-10-21): Coding Plan vs Token Plan vs API aggregator — which subscription shape actually saves money?

Accepted Answer

W43 (10/21-10/26) three-way comparison. **Coding Plan (sells request bundles, priced per seat/month):** Zhipu GLM Coding Pro ¥80/mo and Max ¥120/mo (2000 GLM-5.3-Flash calls + 50 Sonnet 5 calls, riding on GLM-5.3-Flash's 99.4% BFCL-v3 tool-use rate); Kimi Code Moderato ¥99/mo (8000 Kimi K3 calls, 71.8% SWE-bench, 99.7% 1M RAG recall — best for Chinese long-form IDE work); Z.ai GLM Plan ¥122.4/mo (5000 Flash + 1000 GLM-5.2 + 50 Max); internationally Cursor restructured on 10/21 to Free $0 / Plus $8 / Premium $30 (4x) / Super $100 (15x) / Ultra $200 (50x), with Teams $30+/user/mo; GitHub Copilot Pro $10 (+$15 credits +$5 flex) / Pro+ $39 / Max $100 (1 credit = $0.01; Business and Enterprise now contact-sales only).

**Token Plan (sells token buckets / monthly allowances):** Alibaba Bailian Lite ¥9.9 first month / Standard ¥39 (7折 = ¥27.3, unlimited Qwen3.8-Flash at 1M context) / Pro ¥99 (5折 = ¥49.5, all models + 5000 calls); Tencent Hy3 Token Plan ¥150/mo or ¥1100/yr (Hy3 Pro ¥5/¥20 plus 8000 Hy3 Flash calls at ¥0.8/M); Baidu Qianfan ¥9.9/mo (first month ¥4.9, ERNIE-4.5 / 4.5-Turbo).

**API aggregator (pays per token, no seat fee):** TokenRhythm ¥1/¥1 per M unifies DeepSeek V4-Pro / GLM-5.3-Flash / Kimi K3 / Qwen3.8 / Doubao behind one key, with a ¥500 trial credit extended to 10/31. DeepSeek's own valley/peak split (verified 10/21) makes this the only shape that reliably captures off-peak pricing: peak is only Mon-Fri 01:00-04:00 and 06:00-10:00 UTC excluding Chinese public holidays, so off-peak (weekends and holidays included) is roughly half — V4-Pro $0.66/$1.32 input and $1.98/$3.96 output versus peak $1.32/$2.64 and $3.96/$7.92.

**Which one to pick:** under 500 IDE requests per day and a fixed editor workflow → Coding Plan is cheapest (GLM Coding ¥80 or Cursor Plus $8). Stable single-vendor long-context work where you must keep the same model family → Token Plan (Alibaba Standard ¥39 or Tencent Hy3 ¥150). Cross-vendor batch processing, agent orchestration, or anything that should be shifted to valley hours → aggregator (TokenRhythm ¥1/¥1, or DeepSeek directly). The best W43 value is actually using all three at once: one Coding Plan for the IDE, one Token Plan for long documents, and one aggregator account for batch and valley-hour work. Two caveats when comparing across vendors: Claude 4.7+ uses a new tokenizer producing roughly 30% more tokens for the same text, and MiniMax M3 doubles in cost once input crosses 512K tokens (¥2.1/¥8.4 becomes ¥4.2/¥16.8) — normalize both before drawing conclusions.

Question 4

W43 国产大模型横向对比（10/20-10/26）：按价格/上下文/编程/工具调用/RAG 5 个维度选哪个厂商？

Accepted Answer

W43 (10/20-10/26) domestic LLM head-to-head: 8 vendors scored on price / context / SWE-bench / tool-use / RAG. **Top 5 picks by ¥/$ per 1M tokens (1M-context segment):** (1) DeepSeek V4-Pro ¥2/¥8 — 30-round 100% recall + SWE-bench 79.4% domestic #1; (2) DeepSeek V4.1 Flash 谷时 ¥0.25/¥1 (peak ¥1/¥2) — new 1017 build MIT open-source (10/17); (3) GLM-5.3-Flash ¥0.8/¥2.8 — domestic Flash #1 + 99.4% tool-use #1 + 5-折 promo through 10/28; (4) Qwen3.8-Flash ¥1.2/¥4 — domestic 1M Flash #2 + WildClawBench-MM 61.9; (5) Kimi K3 ¥3/¥15 — 1M RAG SOTA 99.7% + 8-折 first-month ¥239 K3 Pro (10/22-11/5). **SWE-bench Verified W43 programming ranking (10/20 updated):** DeepSeek V4-Pro-0813 79.4% (domestic #1, CursorBench 4-sub combined leader) > Qwen3.8-Coder-Plus 76.2% > GLM-5.2 73.1% > Kimi K3 71.8% > mimo-v2-pro 70.1% > MiniMax M3 69.5% > Doubao 2.0 Pro 68.0%. **Tool-use success rate (BFCL-v3 complex nested W43):** GLM-5.3-Flash 99.4% (#1 domestic) > Kimi K3 96.8% > Qwen3.8-Flash 95.2% > DeepSeek V4-Pro 94.5% > Doubao 2.0 Pro 92.1% > MiniMax M3 91.0%. **1M needle-in-haystack RAG (500K doc random 50):** Kimi K3 2.8T 99.7% SOTA > DeepSeek V4-Pro 99.5% > GLM-5.3-Flash 99.1% > Qwen3.8-Flash 98.6% > Doubao 2.0 Pro 97.8% > MiniMax M3 84.6%. **100M token / month cost ranking (cheapest):** V4.1 Flash valley ¥62.5 < GLM-5.3-Flash ¥90 < Doubao 2.0 Pro ¥200 < Qwen3.8-Flash ¥260 < mimo-v2-pro ¥500 < Kimi K3 ¥900 < Hy3 Pro ¥1000 < V4-Pro ¥1600 < MiniMax M3 ¥1650 (vs GPT-5.6 Sol $1400 / Claude Opus 4.8 $77000). **Coding-plan subscription (4 vendor tiers):** Baidu Qwen Lite ¥9.9 / Qwen Standard ¥39 first-month ¥27.3 / Qwen Pro ¥99 first-month 5-折; GLM Coding Max ¥120/month (2000 GLM-5.3-Flash + 50 Sonnet 5); Tencent Hy3 Pro ¥1,100/yr + ¥150/mo; Baidu Qianfan ¥99/mo. **Cross-vendor arbitrage pick:** TokenRhythm 1M aggregate ¥1/¥1 per M (5 models one key) + ¥500 credit to 10/31. **Scenario picks:** (1) Cheapest 1M Pro: V4-Pro; (2) Cheapest batch: V4.1 Flash valley; (3) Best tool-use: GLM-5.3-Flash; (4) Best RAG 1M: Kimi K3; (5) IDE coding: GLM Coding Max ¥120; (6) Multimodal native 1M: Qwen3.8-Omni-Flash; (7) Cross-vendor ¥/$: TokenRhythm 1M. **W43 external risk:** China MoCT 23 AI export controls (10/16) covers Kimi K3 / Qwen 3.8 / Muse Spark Open weights — no API impact, but overseas self-hosting must complete before 2026/11/15. Recommend API-stack for overseas projects.

Question 5

W42 本周免费/优惠 API 汇总（10/13-10/19）有哪些值得申请的硬核福利？

Accepted Answer

W42 free/promo roundup (10/13-10/19). OpenAI GPT-6 Intelligent UI to free ChatGPT (GPT-5.5 retired 10/14); Apple Intelligence China free with iOS 27.3 in December; Anthropic Claude Marketplace 19 partners + METR independent evaluator access; xAI Grok Voice 3.0 $0.08/min trial; Alibaba Qwen Go personal agent free + Qwen Standard first month ¥27.3; OpenRouter 24 :free models (added DeepSeek V4-Pro 1M / GLM-5.3-Flash 1M / Kimi K3 on 10/16) + $10 lifetime credit; 5 open-source LLM launches on 10/15 (Meta Muse Spark Open 816B MoE / NVIDIA Nemotron 3.5 Super VL / Black Forest Labs FLUX.3 Image Klein 6B / IBM Granite Switch 4.1 / Mistral Timestral 209M); OpenAI Rosalind academic API credits; TokenRhythm 1M ¥500 credit extension to 10/31. Stackable hard-core <$30 starting pack: TokenRhythm ¥200 + OpenRouter $10 + Qwen Go free + Doubao work 30-day free + Hugging Face Pro $9 = $200 equivalent value.

AI 太贵？我们帮你省钱

快速推荐

热门模型当前价格

最新价格变动

Coding 方案

按场景选 AI

深度指南

成本计算器

优惠折扣

价格变动历史