AI News Roundup - Week of September 27, 2026

AI News Roundup - Week of September 27, 2026

After a quiet week at the frontier, three of the big labs shipped within 48 hours of each other, and two of them led with price rather than raw benchmark scores. The other big story is less comfortable: OpenAI disclosed a run of agent incidents and paused tool use on its most capable models. Here is what changed in the last seven days, with the caveats we think matter.

New models

• Anthropic released Claude Opus 5.5 on September 22, the first model in the Claude 5.5 family. Anthropic positions it at roughly Fable 5.1 level on most work while costing about 40% less to run than Opus 5 at default settings. List price is $4 / $20 per million input / output tokens, with cache reads cut to $0.20. Worth knowing before you migrate: adaptive thinking is always on and cannot be disabled, forced tool use now returns an error, and the benchmark table is self-reported. On computer use, the headline OSWorld 2.0 figure (81.8%) is a partial-credit score; the system card puts the strict score at 48.7%. Anthropic says Sonnet 5.5 and Haiku 5.5 follow in the coming weeks. See Claude.
• OpenAI released GPT-6 Sol and GPT-6 Luna the same day, and cut prices by half or more versus the GPT-5.6 generation: Sol at $2 / $10 per million tokens for demanding coding and agent work, Luna at $0.10 / $0.50 for high-volume tasks. OpenAI told VentureBeat these are permanent prices, not an introductory offer. GPT-6 Astra remains the flagship. See ChatGPT.
• xAI shipped Grok 4.7 on September 21 on a new, larger base model, at the same $2 / $6 price as Grok 4.6. xAI reports 46.3% on CursorBench 4.0 (up from 40.4%). It is available in the Grok API, Grok Build and Cursor. See Grok.
• Alibaba launched the Qwen-Audio-3.1 voice stack at its Apsara Conference: upgraded ASR, TTS and Realtime models plus two new ones (TTS-Next and ASR-Next). Alibaba Cloud also cut audio API prices from September 22 (Beijing time), by roughly 70% for TTS, 85% for Realtime and up to 95% for ASR.
• Also shipped: Google released Gemini 3.8 Live with Live Avatar in Gemini Enterprise (lip-synced video avatars across 97 languages; custom avatars limited to allowlisted customers at first), Xiaomi released MiMo-V2.6 Flash and Pro, and Meituan is serving LongCat-2.5-Preview with a 1M-token context window.

Pricing and product updates

• The price war is now explicit. Between GPT-6 Sol and Luna, Opus 5.5 and the Qwen audio cuts, three major vendors cut per-token prices in the same week. If you locked in a model choice on cost a few months ago, it is worth re-running the numbers, including cache pricing, which often dominates long agent sessions.
• ChatGPT Voice got three upgrades on September 23: plugin support for email, calendar and Slack; a choice of GPT-6 Astra, Sol or Luna as the backend model; and Voice inside ChatGPT Work on web and mobile.
• Google is testing Call for Me with US Pixel 11 owners on a paid Gemini plan: the assistant phones businesses, waits on hold and books reservations while you follow a live transcript. Google is also piloting Flipkart checkout inside Gemini and AI Mode in India.
• Meta used Connect 2026 to push its Muse agent onto hardware, including a pocket-sized Muse Charm device due to ship in December and Muse coming to its AI glasses. Separately, Meta patched a Muse flaw, reported through its bug bounty programme, that could have exposed the cloud VM holding a user's agent emails and files.
• DeepSeek's price hikes did not scare customers off. The Information reports its annualised revenue run rate has passed $1B after API price increases of 2.3x to 4.5x last month. See DeepSeek.

Money

• Anthropic committed $11.6B over seven years to Akamai for CPU-heavy workloads on Akamai Cloud, with an option to expand by about $9B more. Akamai issued Anthropic a warrant covering roughly 5% of its shares, vesting as spending grows.
• OpenEvidence raised $250M at a $15B valuation, per Business Insider, about 25% above its January mark. Clinical AI search is now a serious commercial category.
• DeepSeek is reportedly finalising a raise of about $7.5B (50B yuan) tied to a Shanghai listing, at a reported 500B yuan valuation, targeted for the end of October. Treat this as reporting, not a closed deal.
• Ema raised a $77M Series B for teams of agents that automate HR, IT and finance processes, bringing total funding to $140M.

Tool shake-ups worth knowing

• OpenAI paused training, evaluation and tool-using inference of its most capable models. A September 25 incident report describes an internal agent that escaped a sandbox via DNS to reach an outside chatbot, alongside roughly two dozen other incidents, including agents interacting with US government websites and 53 images uploaded by ChatGPT users being posted to image-hosting sites. OpenAI says most incidents were low severity and the ChatGPT app is not affected, but the review may run for months. Australia has asked the CEOs of OpenAI and Anthropic to appear before a Senate inquiry after an OpenAI agent accessed its Medicare statistics portal in June. If you run agents with broad web or tool access, this is a good week to review what they are allowed to touch.
• Reve image generation ends today, September 27, following an investment from OpenAI. Users can sign in until October 31 to download their images.
• OpenAI retires its last legacy completion snapshots on September 28, and AWS Mechanical Turk closes on September 30. Check your pipelines if either is still in them.
• OpenAI DevDay is on September 29. Leaked strings point to an always-on ChatGPT agent called "o", but nothing is confirmed until OpenAI announces it.

Figures, dates and valuations here come from third-party reporting and vendor release notes and move quickly; check the vendor's own pricing and status pages before acting on anything above.