Jump within the current audio brief. Story thumbnails mirror the cards below when available.
Same-day briefs for Wed Sep 23, sorted by release time. The selected date stays fixed.
Changing the date preserves the current brief type: AI News AM.
/ai-news/2026-09-23/.Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.
Anthropic released Opus 5.5 on Monday at $4 per million input tokens and $20 output, down 20% from Opus 5, with cache reads slashed 60% to $0.20. It tops Anthropic's own table at 89.9% on SWE-bench Pro and 66.4% on Terminal-Bench 4.0, and ported a legacy HAProxy codebase from C to Rust in 9.5 hours — 51% faster than Fable 5.1. Available now via API, AWS, Google Cloud, and Microsoft.
Ninety minutes after Anthropic's Opus 5.5 landed, OpenAI dropped GPT-6 Sol at $2 in / $10 out and Luna at $0.10 in / $0.50 out — a permanent 50% cut versus the GPT-5.6 line. With Astra at $10/$50, OpenAI now fields a three-tier family spanning enterprise reasoning to high-volume clerical work. The Decoder's read: the prices moved a lot, the benchmarks barely moved at all.
France, holding the September presidency, convened the 15-member Council for a session on AI and international security this morning during UNGA high-level week. Sam Altman attends in person, Dario Amodei virtually, alongside Yoshua Bengio and Hugging Face's Clement Delangue — with Chinese labs DeepSeek and Moonshot invited to make statements. It's the first time the Council has hosted frontier US and Chinese developers together on shared safety concerns.
Four consumers filed a proposed class action in federal court in San Francisco against Anthropic, OpenAI, Google, and SpaceXAI, alleging an illegal agreement to throttle how fast their models improve. The centerpiece is Dario Amodei's September 12 "pace the frontier" essay and the rival-CEO endorsements that followed within the hour. Plaintiffs are seeking treble damages — and the suit lands the same week OpenAI's policy chief said no antitrust waiver is needed for safety coordination.
Xiaomi released its MiMo-V2.6 series Monday, a 1.02-trillion-parameter open-weights flagship that leads open-model rankings and ties Grok 4.7 on the Intelligence Index at a fraction of frontier pricing. But Anthropic named Xiaomi among seven Chinese labs in a September 10 threat report, alleging it routed over 400,000 exchanges through Claude across March and April to harvest training data. Xiaomi has not responded to the allegation.
OpenAI published a position piece arguing for global technical standards governing AI systems that can improve themselves, with the line that "autonomous RSI should only be pursued until it can be done safely." The company also said Tuesday it will open training-phase safety evaluations to third-party groups. The timing — one day before the Security Council briefing — is not subtle.
At ROSCon 2026 in Toronto, Alphabet's Intrinsic released Intrinsic Core on GitHub — collision-free motion planning, perception with NVIDIA FoundationPose support, an on-edge ML inference engine, kinematics solvers, and hardware drivers, all ROS 2 compatible and running locally with no cloud dependency. It ships with a full machine-tending reference design. Apache 2.0, which means commercial use with no strings.
Toyota told investors it expects roughly ¥1 trillion — about $6.4 billion — in annual factory modernization and robotics spend starting in 2028, potentially covering 400,000 robots across its plants, group companies, and major suppliers. The figure spans humanoid and non-humanoid systems, industrial arms, and automated logistics. Toyota has not committed to the full number or how long spending would run at that level.
A crowdsourced contest run by Terence Tao's SAIR foundation catalogued all 25,000 polynomial-to-Galois-group correspondences acting on 24 roots, closing the last open case for the Mathieu group M23. The win came from a German team that blended professional mathematics, amateur participation, brute computation, and AI-assisted search. Notably, the model didn't do it alone — and the winning team used AI mostly to write an upload script.
OpenAI announced Tuesday that Patreon co-founder Sam Yam will lead a newly formed Creator Product division, building tools aimed squarely at content creators. Yam's public framing leaned on human creativity staying primary with AI as the instrument. It's a notable hire given how much friction OpenAI has absorbed from creative industries over the past two years.
