Audit finds Anthropic's AI improving about 1.9× faster, likely mostly thanks to AI itself
After 11 weeks inside Anthropic, METR estimated its models improve about twice as fast as before, probably mostly thanks to AI's own help.
METR, the independent evaluator behind the “task length” charts on this site, spent eleven weeks (August–October 2026) inside Anthropic. It saw internal test scores, compute figures, staff surveys and interviews. It also looked at unreleased models it calls “Model 3” and “Model 4”.
On Anthropic’s internal capability index (one score combining its tests), progress since October 2025 ran about 1.9× faster than from March 2024 to October 2025. That is roughly two years of progress in one (95% CI 1.5–2.8×). Model 4 scores above 50% on all 15 tests used to rate it. Nine months earlier, Claude Opus 4.6 was below 50% on many of them.
METR is confident the speed-up did not come from faster growth in training compute (the computing power used to train models). It guesses that most of it, perhaps 80%, comes from AI helping with AI research. It found no clear bottleneck for the next 12 months and warns of “a significant risk” that progress speeds up further.
Anthropic’s safety policy treats AI doubling the speed of its research as an early-warning line. Anthropic’s public August report said the line had not been crossed. METR calculates that Anthropic’s own private figures imply 1.97×, and says it is plausible the line was crossed by July 2026. METR still agreed that the direct risk of catastrophe at that point was low.
What it means
AI speeding up its own development is the key step in forecasts like AI 2027, and this is the first independent measurement from inside a lab suggesting it is happening at scale. It covers one company and its own tests, and does not mean AI now does the research alone. It does make faster change over the next year a real possibility, which METR flags as a risk.
Caveats
- Anthropic redacted most of the evidence METR found most informative, so readers largely have to rely on METR’s judgement.
- The index measures well-defined, automatically graded tasks. METR warns the speed-up may not map cleanly onto real-world impact.
- Anthropic was behind the industry’s best before late 2025, which may inflate its measured speed-up. Including other labs’ models gives about 1.4×.
- “Model 3” and “Model 4” are METR’s labels for unreleased models, not product names. METR says it is not confident in its bottleneck assessment.