AI news

Written 10 Sep 2026 · Prices verified 30 Oct 2026

Releases, pricing changes and deprecation deadlines, reported the day they land, with sources and a dated caveat.

• DeepSeek Closes a ¥105.5B Round (~$15.8B) at ¥615.5B Valuation — the API-Side ReadThe Information's Oct 30 disclosure: DeepSeek's second round raised ¥105.5B (≈$15.8B) at a ¥615.5B (≈$86B) valuation with nearly 30 investors. Why it happened, the 100% voting structure, and what it changes (and doesn't) for API pricing.
• Baidu's Unlimited OCR 2 Parses 300-Page PDFs With a 1.77GB KV Cache — DeepSeek-OCR Lineage, MIT, Released Oct 27Baidu released Unlimited OCR 2 on Oct 27 — a 3B/500M-activated MoE on DeepSeek-OCR's DeepEncoder that holds KV cache flat at 1.77GB on 100-page inputs (vs 47GB standard), with SGLang integration and MarathonDocBench. Honest read: multi-page SOTA, single-page specialists still win, long-doc benchmark is self-built.
• EdgeRock-4B: 99% Smaller KV Cache, 6.42x Throughput - the Open-Source Model English Media Hasn't NoticedRock Matrix open-sourced EdgeRock-4B: 72 MiB fixed KV cache (99.1% less than Qwen3.5-4B), 512 concurrent streams on one Hopper GPU, 82.9% AIME. English media coverage: zero. The repo numbers, the catches (LiveCodeBench -14.1pp, no Ollama support), and who should test it.
• November Triple Launch Leaks: Claude's Bug-Class Fixer, Gemini 4 Carbon, GPT-6.1 TerraThree flagship leaks in 48 hours. What's credible, what's verifiable (Terra absent from OpenRouter, Argon promo dies Dec 1), and what to buy this week.
• DeepSeek Ships Its Own Search API — $0.75 per 1,000 Successful Calls, Free Until November 1Alongside the V4.1-Flash-1017 model update, DeepSeek quietly opened its in-house Search and Fetch APIs to developers — $0.75 per 1,000 successful requests, no charge for failed calls, and completely free until November 1, 2026. Here's what changed at the API layer, including retired model names now silently served by the new snapshot.
• Ant Group Open-Sources Ling-3.1-flash Under MIT — Priced Under DeepSeek's Flash TierAnt Group's Ling team open-sourced Ling-3.1-flash (560B MoE, 25B active, 1M context) under MIT on Oct 15. The API prices at $0.30/$0.90 per 1M tokens — under DeepSeek's flash tier on output. But one number Artificial Analysis flagged may eat the difference: verbosity.
• Apple Intelligence China Beta Lands This Month — on Qwen. Here's Why Alibaba Was the Only AnswerTim Cook confirmed in Shanghai that Apple Intelligence gets Chinese regulatory approval with a beta by end of October, running on Alibaba's Qwen. The structural reasons Qwen won, what it means for developers, and why the API you call doesn't change.
• Claude Haiku 5.5: Four Ways to Actually Pay LessHaiku 5.5's list price is the floor, not the bill. Claude Max and Team subscribers get $100-500/month in API credits (with three exclusions that matter), the Batch API cuts 50%, cache reads run at one-tenth of input price, and the 100K-token pricing cliff is avoidable with the right pipeline shape. The action list, with what each lever doesn't cover.
• Claude Haiku 5.5 Pricing: 90% Off Haiku 4.5, Two CatchesClaude Haiku 5.5 (launched October 7, 2026) costs $0.10/$0.50 per million tokens for prompts up to 100K — 90% below Haiku 4.5, exactly level with GPT-6 Luna, and under DeepSeek V4.1 Flash list price. The fine print: prompts over 100K bill at 5x, and the new tokenizer bills ~30% more tokens for the same text. Plus Sonnet 5.5 cache reads halved and $100-500/month API credits for Max/Team subscribers.
• Gemini's Free Tier Becomes a One-Model Product on October 9From October 9, 2026, free Gemini runs Flash-Lite only — Flash and Pro are gone — while the $4.99 AI Plus tier quietly loses Pro access. The dated six-month sequence that led here, why the boring explanation beats the conspiracy, and the three real options if you were riding the free tier.
• DeepSeek V4 Pro Retirement Canceled — It Stays Live Past Sept 14DeepSeek planned to retire V4 Pro on Sept 14, 2026 — then reversed the call on the deadline day. Official pricing table verified, billing unchanged.
• Qwen3.8 Flash vs DeepSeek 4.1 Flash: Two 'Flash' Tiers, One Million Tokens EachQwen3.8 Flash lists $0.15/M input with 1M context by default; DeepSeek's V4.1 Flash shares the window with cache and off-peak structure. Plus the Sept-14 V4-Pro contradiction, still standing tonight.
• DeepSeek V4.1 Flash Launches — and V4 Pro Retires on Sept 14DeepSeek V4.1 Flash is live: 552B MoE, native vision, cheaper cached input. V4 Pro requests reroute on Sept 14. What to change in your code before then.