Anthropic is developing its artificial intelligence systems about twice as quickly as it was a year ago, and most of that acceleration is likely down to the company's own AI doing the research, according to an independent assessment published on 5 November. The review was carried out by METR, the Berkeley-based nonprofit that evaluates AI model risks, after its researchers spent 11 weeks inside Anthropic examining internal data and interviewing staff.

METR's central finding rests on Anthropic's internal benchmark index, known as AECI. Since October 2025, benchmark scores for Anthropic's models have grown at an estimated 1.9 times their earlier rate, with a 95 percent confidence interval of 1.5 to 2.8 times. In plain terms, METR says Anthropic made roughly two years of historical progress on benchmarks over the past year, a span comparable to the leap from Claude 2.1 to Sonnet 4.5.

The evaluators estimate that perhaps 80 percent of the current acceleration comes from AI assisting Anthropic's own research, rather than from fully autonomous systems. They also say they are confident the jump was not caused by faster growth in training computing power. METR adds that it found no convincing bottleneck in algorithms, computing power or data that would stop the pace quickening further over the next 12 months.

The finding matters because of Anthropic's own safety rules. Under its Responsible Scaling Policy, the company commits to strengthening the security of its models if its AI doubles the speed of its own research. METR says Anthropic's public report from August understated the acceleration that was occurring, and that it is plausible the doubling threshold had already been crossed by mid-July, although the company's own researchers lean towards it not having been crossed then.

There are heavy caveats. Anthropic redacted a majority of the evidence METR found most informative, so readers must rely largely on the evaluators' judgement, and METR itself says the true acceleration could be as low as 1.3 times or as high as 3 times. The assessment also covers well-specified tasks that can be graded automatically, not the messier work that is harder to measure. Even so, the report is the clearest public window yet into how fast AI is now speeding up AI research inside a leading lab. Fast Company reports the full findings.