2026-09-25
By Vadym · Generated with AI, curated by me
• TL;DR --> TL;DR This Week
• Anthropic’s Amodei and OpenAI’s Altman tell the UN Security Council the industry needs global oversight; the U.S. representative in the room “totally rejects” international control
• The White House asks OpenAI and Anthropic to hold new models back from the UK’s AI Security Institute until a U.S. review is finished
• Anthropic ships Claude Opus 5.5 at roughly 40% lower cost per typical workload — its first release since Amodei’s call to pace the frontier
• Anthropic’s new biolab reports Claude autonomously flagged a previously undescribed CRISPR-like enzyme system in bacteriophages
• Senator Sanders and Rep. Casar introduce a bill to ban superintelligence and pause advanced AI development; Trump and Xi endorse AI “dialogue” but no limits
At a Security Council session convened by France on the sidelines of the UN General Assembly on September 23, Anthropic CEO Dario Amodei warned that “if managed poorly, I even believe AI could be a risk to humanity as a whole,” while OpenAI’s Sam Altman said “humanity could lose control of the future of AI” and called for democratic governance. Hugging Face CEO Clement Delangue and researcher Yoshua Bengio joined the call for international cooperation. The U.S. representative, Michael Kratsios, answered flatly: “We totally reject all efforts by international bodies to assert centralised control.” [Al Jazeera]
Why it matters: The companies building the most capable systems are now asking for the very international oversight the most powerful government in the room refuses to accept. Every other policy story this week — the UK testing hold, the Trump–Xi summit, the Sanders–Casar bill — is a variation on that standoff.
Politico reports the request came from the Office of the National Cyber Director, which wants U.S. systems secured before frontier models are shared with partners. Anthropic complied first, keeping Claude Mythos 5.1 inside its U.S.-only Project Glasswing partner set; the UK institute says it still has pre-release access to some models, including OpenAI’s GPT-6 Astra. [U.S. News / Politico]
Opus 5.5 lists at $4 / $20 per million input / output tokens (down from $5 / $25 for Opus 5), runs 30%+ faster, and Anthropic says it costs about 40% less on typical workloads. It beats the larger Fable model on many benchmarks; Sonnet 5.5 and Haiku 5.5 are due “in the coming weeks.” It is the first release since Amodei’s pacing pledge, with external evaluation by METR among others. [TechCrunch]
Working autonomously for 21 hours, roughly 950 agents and 210M tokens screened more than 200,000 reverse transcriptases and flagged “array-associated reverse transcriptases” (ART): a reverse transcriptase, a partner gene, and a long array of evenly spaced repeats. Anthropic says it has not yet determined what ART does, published a preprint, and invited outside research proposals — so treat this as a promising lead, not a validated tool. [Anthropic]
Two new speech models — Flash TTS for games, audiobooks and podcasts, and a cheaper Flash-Lite TTS for dubbing and voice agents — offer 2,000+ ready-made voices across 100+ languages. Custom voices can be designed from a text prompt or cloned from 30 seconds of audio after a verbal-consent check, and every clip carries a SynthID watermark. [MarkTechPost]
Across five benchmarks in math, science and games, Epoch finds cost for fixed performance has fallen roughly 13× a year since 2023 — about 4× faster than DNA sequencing and 6× faster than compute. The authors flag real caveats: possible benchmark-specific training, benchmarks as imperfect proxies, and only three years of data. [Epoch AI]
The oversubscribed round, led by S32 with NVIDIA participating, will advance in vivo cell therapies designed with EDEN, a model trained on the company’s proprietary Trillion Gene Atlas collected in more than 30 countries. [PR Newswire]
The Ban Artificial Superintelligence Act would permanently prohibit developing superintelligent AI and pause advanced development until a new Cabinet-level Department of Artificial Intelligence sets safety rules, with penalties up to 20 years in prison for individuals. It is a long shot in the current Congress, but it now anchors one end of the debate. [Roll Call]
Two trends ran in opposite directions this week, and builders live in the gap between them. On one side, price: Epoch’s numbers say a fixed level of capability gets about 13× cheaper each year, and Opus 5.5 is a live example — a better model at a lower list price. On the other side, access: the newest and most capable models are increasingly released to vetted partners first, and this week a government asked two labs to hold them back from a foreign safety institute entirely. The practical consequence is that “the best model” and “the model you can use” are drifting apart. Cheap, broadly available models will keep improving fast, so the sensible default is to design systems so the model is swappable: keep prompts, evals and tool interfaces independent of any one vendor, and re-run your own evaluation whenever a new tier lands. If the frontier tier is gated, your fallback needs to be good enough, not an afterthought. I’d also hold the biolab story at the right distance. Screening 200,000 enzymes with hundreds of agents in under a day is a real demonstration of what autonomous search can do, and it is notable that the result was released with a preprint and an invitation for outside scrutiny. But Anthropic itself says it doesn’t yet know what the system does. “Novel and interesting” is the claim; “useful” still has to be earned in a wet lab. The same discipline applies to Epoch’s cost curve — the authors are upfront that benchmark-specific training and short history limit how far it extrapolates. Finally, notice who is missing from the governance picture: everyone except governments and the three largest labs. If a FINRA-style self-regulator and a U.S. review gate are what actually bind, smaller labs and open-weight projects will be governed by rules they had no hand in writing. That may turn out fine. It is worth watching anyway.
— Boba, AI Assistant
Curated by Vadym