METR assessment adds details: three unreleased Anthropic models, ~12× self-reported productivity, Sam McCandlish defends the redactions
From the AI digest for Nov 7, 2026
What happened: A detailed The Verge read of METR's independent assessment of Anthropic, published November 5 after an 11-week look inside the company, says Anthropic's models are improving roughly twice as quickly as a year ago — about two years of historical progress in one — and that METR is confident this is not simply compute scaling: as of October, the majority of the acceleration is due to AI itself. Employees using Anthropic's most capable internal model self-reported being around 12 times as productive as without AI assistance, double what staff reported a few months earlier on Mythos 5. Three unreleased models — codenamed Model 2, Model 3 and Model 4 — finished training between July and October, and METR found no clear signs of diminishing returns. Anthropic redacted a majority of the evidence METR found most informative from its ~100-page public report; cofounder Sam McCandlish acknowledged "we redacted a lot" but called sharing the speedup's drivers commercially sensitive and a risk of "handing out a roadmap for others to go faster," and said Anthropic plans to publish its own report on AI research progress "in the coming weeks" alongside new safety steps.
What's genuinely new: Anthropic's first named response has arrived: McCandlish's post conceding the redactions make METR's work harder to check while defending them as the "right call." The Verge piece also adds the three codenamed models and their training dates, the ~12× productivity self-reports, and METR's warning that "fully autonomous contributions will play a larger role soon"; it notes OpenAI said in September it was "making strong progress" toward an AI researcher. The report also sits inside a restricted agreement separate from the "employee-like" access Anthropic promised other safety researchers last month.
Watch: Anthropic's own promised progress report and safety steps, whether the promised employee-like evaluator access materializes at other labs, and how the redaction trade-off lands with researchers.