Key facts
- METR said Anthropic finished training three models not yet available to the public over the prior six months.
- METR said Anthropic redacted a majority of the evidence it found most informative for its risk conclusions.
- Anthropic staff self-reported an average productivity increase of roughly 12x for the most capable internal model as of October, per METR.
- METR said it considers the Slack poll data weak evidence and that respondents may not have thought carefully about answers.
- METR estimated Anthropic made roughly two years of historical progress over the last year, primarily via AI assistance.
- Jacob Coxon, an ex-Anthropic researcher who quit in September, said the companies are racing to self-improving superintelligence.
Safety researchers embedded inside Anthropic's San Francisco headquarters released a report on November 5, 2026 about the company's AI development. METR, a research group, said it received "unprecedented access" inside the company for nearly three months, with six researchers interviewing employees, monitoring internal Slack channels and examining unreleased models on company hardware.
METR said Anthropic redacted a majority of the evidence it found most informative for its risk conclusions. "While we were able to freely share our highest-level conclusions about risk, Anthropic redacted a majority of the evidence we found most informative for these conclusions," METR said in the report.
The report said Anthropic had finished training three models not yet available to the public over the prior six months, including one from 2022 and others finished in July, in August or September, and in September or October. METR said those internal-only models help Anthropic improve its capabilities faster, and that keeping state-of-the-art models private gives product and research teams a competitive advantage.
Anthropic turned over internal Slack poll data showing staff reporting productivity gains. METR wrote that the average self-reported increase was roughly 12x for the most capable internal model as of October, but cautioned that it considers Slack polls weak evidence because they may not represent all employees and respondents may not have given answers much thought. One researcher told METR that work was "Now only bottlenecked on ideas & compute, no longer bottlenecked on code."
METR said it did not see observed bottlenecks that would prevent acceleration from increasing further over the next year. "The lack of observed bottlenecks or strong diminishing returns means there is a significant risk that further capabilities progress could lead to even faster acceleration, though, as we say below, we may be missing important bottlenecks," it wrote.
METR found that Anthropic engineers are increasingly delegating oversight and training to other AI models, a process known as recursive self-improvement, and estimated the company made roughly two years of historical progress over the last year. Researchers said they were told the acceleration came "primarily via AI assistance rather than fully autonomous contributions," with signs fully autonomous contributions will play a larger role soon.