Anthropic AI improving itself? METR finds progress 1.9x faster with AI help
An 11-week outside check by safety group METR says Anthropic's models now improve 1.9 times faster than before, with AI doing perhaps 80% of that speed-up.
Anthropic’s AI is getting better almost twice as fast as it used to, and much of that speed now comes from AI doing the work. That is the finding of METR, a US nonprofit that tests advanced AI, in a report published on Thursday, 5 November after an 11-week look inside the company.
It is one of the first times an outside group has been allowed to measure, with a lab’s own internal data, how far AI is speeding up the building of AI.
What did METR find?
- 1.9x faster: since October 2025, Anthropic’s models have been gaining capability about 1.9 times faster than the earlier trend. METR’s statistical range runs from 1.5x to 2.8x.
- AI is most of the reason: METR estimates that “perhaps” 80% of the current speed-up comes from AI tools helping Anthropic’s researchers and engineers. It flags that number as very uncertain.
- Not just bigger computers: METR said it was “confident” the jump was not caused by faster growth in the computing power used for training.
- No wall in sight: it found no convincing bottleneck (in algorithms, computing power or data) likely to slow things down over the next 12 months.
The measure METR used is Anthropic’s internal version of a capabilities index built by the research group Epoch AI, which combines many benchmark scores into one line over time.
How the check worked
The review ran from 12 August to 29 October. METR says it saw internal benchmark scores, training data figures, headcount and code statistics, and interviewed about 15 Anthropic staff. It did not test unreleased models itself, and it says it took no payment from Anthropic, though the company supplied free usage and devices.
There are limits. Anthropic redacted much of the most detailed evidence, saying that information could help rival labs. Staff polls inside the company claimed AI had made them up to 12 times more productive; METR called that “very loose evidence” and did not rely on it.
Is this AI improving itself?
Partly, and that is why the report is getting attention. If each new model helps build the next one faster, progress can compound. METR’s caveat matters, though: AI is “not yet able” to make highly significant, fully independent contributions to the most important parts of training a model. Humans are still steering.
The timing is awkward for Anthropic, which is preparing a stock market listing as investors argue over its valuation. The company did not immediately comment on the report.
Why it matters in Pakistan
Faster AI progress lands first on the work our freelancers and IT firms sell: coding, testing, data work and writing. If tools improve every few months instead of every year, skills go stale faster too. Students and developers who learn to work with AI agents, rather than compete with them, are the ones best placed.
METR says it wants other labs to accept the same kind of outside check. Whether OpenAI or Google agree is the next thing to watch.
