METR report says AI is accelerating Anthropic’s model development
publishers 182articles 359first reported 24 Sep, 16:43 UTCdeveloping since 24 Sep · 52 editions
Anthropic’s use of artificial intelligence in its own research has helped its models improve at roughly twice their previous pace, according to a report published Thursday by the independent AI safety group METR.
The report followed an 11-week investigation into Anthropic’s internal research and development. Benchmark results indicated the company made about two years of historical progress over the past year. METR estimated that since October 2025, model scores had grown at 1.9 times their earlier rate, with a 95% confidence interval of 1.5 to 2.8 times.
METR said the acceleration was not primarily caused by increased computing power. It attributed most of the faster progress as of October 2026 to AI assistance, supported by employee interviews and informal internal polls. Anthropic employees using the company’s most capable internal model reported productivity averaging about 12 times what they expected without AI, The Verge reported.
Human researchers remain in control, with models largely serving as research assistants, according to The Indian Express. However, METR said there were signs that fully autonomous contributions could soon play a larger role. Recursive self-improvement would involve an AI system autonomously designing and developing a successor without human intervention.
Separately, Anthropic engineer Addy Osmani warned that extensive reliance on coding agents can lead to “cognitive surrender,” in which users trust AI output without retaining ownership or understanding. He urged engineers to maintain verification practices and continue exercising skills they do not want to lose.
HOW THIS STORY WAS MADE
Written from 8 articles, 3 with the publisher's own text; 8 independent newsrooms once syndicated copies count as one; this version written 930 h after the record first saw the story; the editor kept it as written.
- Publishers
- 182
- Source articles
- 359
- Given to the writer
- 8, 3 with the article's own text
- Independent newsrooms
- 8
- This version written
- 930 h after the record first saw the story
- Second model (editor)
- kept as written
- Publication gate
- passed
- Human review
- none
WHAT THE EDITOR DID
The second model checked the writer's draft against the same sources and published it as written.
Written by a language model from the sources above, then checked by a second model that may only cut, attribute or correct. How it works →