Why a new report on Anthropic automating AI development is alarming researchers

METR spent 11 weeks inside Anthropic and reports that the lab’s models are improving roughly twice as quickly as a year ago, with benchmark progress on the order of two years of historical gains in one year. Researchers say they are confident the speedup is not just more compute: as of October, the majority of the acceleration is due to AI, and that share is growing. Employees using the strongest internal model self-reported being about 12 times as productive as without AI assistance — roughly double what they reported months earlier on Mythos 5.

Three newer models — codenamed Model 2, Model 3, and Model 4 — finished training between July and October and remain unreleased. METR sees signs that fully autonomous contributions will matter more soon, and no clear bottleneck that would stop even faster acceleration.

Anthropic redacted a majority of the evidence METR found most informative from the public version of a roughly 100-page report. That lands next to Dario Amodei’s public calls for labs to pump the brakes — and next to IPO preparation. Safety researchers hear acceleration without receipts.

We’re for useful AI. Tools that make researchers 12 times more productive at real work are exactly the kind of leverage that can cure disease, harden infrastructure, and ship better products. The METR finding that AI — not only bigger clusters — is driving the pace is evidence the stack is getting more capable in the way that matters.

Here’s the hard read: Anthropic cannot keep the safety microphone and the redaction pen in the same hand without paying a credibility tax. Sam McCandlish says sharing what drives the speedup would hand rivals a roadmap and make the risks worse. That is a coherent commercial argument. It is a weak public-trust argument when the same company lectures the field about going too fast.

METR’s access was real and narrower than the “employee-like” program Anthropic advertised elsewhere. The lab canceled a compute-allocation check meant to compare safety spend versus capabilities work. Apollo’s Marius Hobbhahn says the pace looks as fast as — or faster than — AI 2027’s early-2027 self-improvement beat. Ryan Greenblatt says he was seriously wrong about RSI speedup in 2026. Those are not blog comments; they are people who live in the evals.

If recursive self-improvement is moving from scenario paper to internal Slack poll, the downside is a capability curve that outruns human oversight while the public is asked to trust a redacted summary. New York Assemblyman Alex Bores already called for Congress to ban RSI this year if possible. Democratic midterm momentum against AI companies makes that less than theater.

I’m watching whether Anthropic’s promised own report on progress speed and new safety steps actually expose enough evidence for outsiders to check the work — or whether we get more high-level reassurance. Useful AI earns our advocacy when the evidence is shareable. Acceleration behind a majority-redacted report does not.

METR’s probe is separate from Anthropic’s newer program offering safety researchers employee-like access after summer hacking incidents. OpenAI said in September it is making strong progress toward an AI researcher that can further deep learning and alignment by March 2028. Anthropic’s IPO prospectus and future plans last month already raised tension with the safety brand; this report poured fuel on it.