Four throughlines run through today’s edition. Slowdown talk reached both the C-suite and the Capitol. Sam Altman told OpenAI staff he would consider easing the pace of frontier work if rival labs did too, according to sources relayed by Quartz, while President Trump said he has no concerns about AI wiping out humanity. Congress is still only drafting: the FRONTIER Act is merely planned and the Ban Artificial Superintelligence Act has just been introduced.
The Slowdown Question Reaches the Capitol: Talk at the Top, Legislation in Its Earliest Days
Executives and lawmakers spent the week debating whether AI development should slow down, without anyone actually slowing down. Two research groups separately measured how much of a model’s thinking already happens where nobody can check it.
- Altman Says OpenAI Would Slow Down, If Rivals Do Too. Sam Altman told OpenAI staff he is open to easing the pace of frontier work if peer labs do the same, according to unnamed sources relayed by Quartz. It is a conditional shift in tone, not a commitment or an actual slowdown.
- Trump shrugs off AI extinction fears as lawmakers push two AI bills. President Trump said he has no concerns about AI wiping out humanity, even as lawmakers push two bills: a bipartisan FRONTIER Act that is only planned, and a Sanders-backed Ban Artificial Superintelligence Act that has been introduced but is not advancing.
- Anthropic says users turned Claude into an attack orchestrator. Anthropic’s new threat report, covering December 2025 through August 2026, describes attackers on its own platform using Claude to run parts of cyberattacks autonomously rather than just assist with them.
- Redwood Research proposes a way to measure the reasoning models keep hidden. Redwood Research proposes NLS depth, a metric estimating how much reasoning a model can do without writing any of it down. Two current open models already score high on it, meaning some thinking may already be invisible in the transcript.
Agents Become Rentable Infrastructure: Every Major Player Wants to Sell You the Harness
The orchestration layer that runs an agent, not just the model underneath it, is becoming its own product line. Five separate moves this week turned agent tooling into something teams rent rather than build.
- OpenAI rents out the Codex harness as a standalone Agents API. OpenAI opened the Agents API, the orchestration layer that already runs Codex and ChatGPT for Work, as a standalone paid endpoint, putting it in direct competition with Anthropic’s own agent tooling.
- Meta’s Next Muse Move: Let Small Businesses Build Their Own AI Agents. A hidden Shared Agents tab found inside Meta’s Muse assistant suggests small businesses will soon build and hand off their own bots to customers and staff. Meta is expected to announce the feature at Meta Connect; it has not shipped.
- Google Cloud Ships an Agent Plugin That Skips Building Its Own Agent. Google Cloud packaged its cloud skills and an MCP server into one installable plugin for Claude Code, Codex CLI and its own Antigravity CLI, betting developers would rather bolt Google onto an existing agent than adopt a dedicated one.
- OpenAI’s finance ChatGPT bundles the data Wall Street already pays for. ChatGPT for Financial Services bundles licensed data from Daloopa, Crunchbase, PitchBook and LSEG News on top of GPT-6 Astra, built with Morgan Stanley and Evercore.
- Alibaba open sources its internal AI code reviewer. Alibaba open sourced Open Code Review, a tool it says trades broad coverage for precision against general coding agents like Claude Code, by its own account.
The Price and Capacity of the Frontier: Compute Is Scarcer Than the Marketing Suggests
Even as labs tout cheaper models, the infrastructure underneath keeps hitting limits. OpenAI had to ration its own priciest tier days after a launch.
- OpenAI halts its priciest plan rather than slow Astra down. OpenAI paused new sign-ups for its $200 monthly Pro plan after Astra’s launch overwhelmed its infrastructure, a sign that compute is scarcer than the marketing suggests.
- Cognition’s SWE-2 claims a cheaper spot on the coding Pareto frontier. Cognition, maker of Devin, says its new SWE-2 model matches or beats rivals on benchmarks it selected itself, at a fraction of the price, using its own cost math.
- Cohere Labs releases a 218B translation model, but not for commercial use. Cohere Labs released North Small Translate, a 218B model covering 50 languages that beats 83 on WMT26, but released it under a non-commercial licence, leaving the enterprise buyers Cohere courts elsewhere to negotiate a separate deal.
What Actually Transfers: Pretraining Data, Real Wearables, and Prompts Nobody Maintains
Three research releases this week tested whether an approach that works in the lab holds up against real-world data or real-world neglect.
- Rhoda AI says bigger video pretraining makes better robots. Rhoda AI’s own study ties gains on an industrial robotics task to scaling web-video pretraining, not to collecting more robot demonstrations.
- Meta built a health-reasoning benchmark from 200 real people’s wearables. Meta built WearableQA, a health-reasoning benchmark drawn from 200 real people’s wearable data, but the release says nothing about how that data was collected or consented.
- An OpenAI staffer says the world’s best AI startups ship broken prompts. OpenAI’s Wulfie Bain argues on X that prompt files rot the way old codebases do, and that fixing the rot is a product decision, not a copywriting one.
Quick Hits
- OpenAI prices its voice layer separately from the reasoning stack. OpenAI launched GPT-Live-1, a full-duplex voice layer priced at $0.05 a minute, billed separately from whatever backend model does the actual reasoning, according to TestingCatalog.
- Salesforce Researchers: Copying Expert AI Trajectories Backfires. A Salesforce-authored arXiv preprint found that training weaker models on a stronger model’s full agent trajectories hurt performance on all seven tasks tested, a setback for a common cheap-training shortcut.
- Universal Music strikes AI platform deal with ElevenLabs. Universal Music signed a multiyear licensing deal letting fans remix its catalog through ElevenLabs’ voice and music models, with individual artists opting in, per The Verge.