METR says AI is now driving most of Anthropic’s speedup, and the evidence is mostly redacted

The Verge reports that METR spent 11 weeks inside Anthropic. Its finding: the lab’s models are improving roughly twice as quickly as a year ago, with benchmark progress equal to about two years of historical gains in one year. The researchers say they are confident the speedup is not just more compute. As of October, most 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 help, 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 soon matter more, and no clear bottleneck that would stop even faster acceleration.

Anthropic removed a majority of the evidence METR found most informative from the public version of the roughly 100-page report. 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 from a company whose CEO, Dario Amodei, publicly calls for labs to slow down, and which is preparing for an IPO whose prospectus already raised questions about its safety brand (Anthropic’s compute economics).

METR’s access was real but narrower than the “employee-like” program Anthropic has advertised for safety researchers after the summer hacking incidents (that week’s coverage). The lab also cancelled a compute-allocation check meant to compare safety spending with capabilities work. Holding the safety microphone and the redaction pen at the same time carries a credibility cost.

The reactions come from people who work on evaluations. Apollo’s Marius Hobbhahn says the pace looks as fast as, or faster than, the early-2027 self-improvement step in the AI 2027 scenario. Ryan Greenblatt says he was seriously wrong about recursive self-improvement speedup in 2026. Their concern echoes warnings from inside Anthropic itself, including a researcher who resigned over self-improving AI (our coverage) and safety leaders discussing extinction risk (coverage).

Anthropic is not alone in automating research. OpenAI said in September it is making strong progress toward an AI researcher that can advance deep learning and alignment by March 2028. A model that writes the tools to make itself better is the scenario Anthropic’s own models were released alongside (Mythos and Fable coverage).

Tools that make researchers 12 times more productive at real work are the kind of leverage that can help cure disease, harden infrastructure and ship better products. METR’s finding that AI, not just bigger clusters, is driving the pace shows the stack is improving in a way that matters.

If recursive self-improvement is moving from scenario paper to internal survey, the risk is a capability curve that outruns human oversight while the public is asked to trust a redacted summary. New York Assemblyman Alex Bores has called for Congress to ban RSI this year if possible, and Democratic momentum against AI companies in the midterms makes that more than talk. Anthropic has promised its own report on progress speed and new safety steps. Whether it shows enough evidence for outsiders to check is the test. Useful AI earns support when the evidence can be shared. More in our Anthropic coverage and research coverage.