Outside check finds AI is roughly doubling Anthropic's pace of progress

What did METR find at Anthropic?

METR, a nonprofit that tests AI systems for risks, published its report on Thursday, November 5, 2026. Its staff worked inside Anthropic from August 12 to October 29, with access to internal benchmark scores, training data, staff surveys and about 15 interviews with key researchers.

METR's main conclusion is that Anthropic is likely improving its models on well-defined tasks at around twice the rate it managed between March 2024 and October 2025. METR calls that estimate highly uncertain and says the true figure could be closer to 1.3x or 3x.

It also says a majority of that speedup likely comes from Anthropic's heavy use of AI in its own research. In METR's own rough wording, that's "a majority (perhaps ≈80%)" of the current speedup, a guess it makes with "substantial uncertainty."

How did METR measure the speedup?

The main yardstick was AECI, Anthropic's internal version of the Epoch Capabilities Index, a score that rolls many AI test results into one number. Since October 2025, Anthropic's AECI has grown about 1.9 times faster than before, with a 95% confidence range of 1.5x to 2.8x. METR says that works out to roughly two years of earlier progress packed into the last year.

METR is "confident" the jump didn't come from faster growth in training compute, the raw computing power used to build each model, which has driven a lot of AI progress in the past. Swapping in other reasonable ways to calculate the score gave answers from 1.1x to 2.9x.

The data includes three Anthropic models that haven't been released. METR calls them Model 2, Model 3 and Model 4, and says they finished training in July, August to September, and September to October 2026.

How much of the speedup comes from AI?

Staff answers point the same way, though METR treats them carefully. In informal Slack polls, people using Anthropic's best internal model reported being about 12 times as productive on average in October, as The Verge highlighted. METR calls those polls "only very loose evidence."

Interviews with researchers centered on roughly a 2x speedup in their area of work. Most of the help today comes from AI assisting human researchers, not AI working alone. METR adds that there are "signs of life" for fully autonomous AI contributions soon.

METR also saw no convincing bottlenecks, such as running out of data or computing power, that would slow progress in the next 12 months. Its view changed several times during the work, so it isn't confident about that.

What does the report leave out?

A lot. Anthropic had the right to redact non-public information, and METR says Anthropic "redacted a majority of the evidence we found most informative." METR says that leaves the public relying on its judgment "much more than would be ideal."

Anthropic also cut a planned part of the review that would have checked how much computing power it spends on safety versus building more capable models. METR says the benchmark could overstate progress too, since it mostly measures clean, easy-to-grade tasks.

Anthropic cofounder Sam McCandlish defended the redactions on X, according to The Verge, saying that publishing exactly what drives the speedup could hand rivals "a roadmap." The report lands as Anthropic prepares to list its shares.

What it means if you use Claude or other AI tools

If AI companies really are improving their models twice as fast, the tools people use every day will keep changing quickly. New models, features and plan changes may arrive more often, and what worked last quarter can get outdated sooner.

The report also adds fuel to the safety debate. An outside group with inside access is saying AI now helps build the next AI, and The Verge reports that the finding has alarmed a number of safety researchers.

X reactions

METR shared its findings in a thread on X. Two of the key posts:

3. Interviews and surveys with staff are consistent with AI usage driving current acceleration, though this evidence is very hard to interpret. Staff self-report dramatic productivity gains, but say the most important AI R&D contributions aren’t yet fully automated. https://t.co/B2pxp0jpLZ

— METR (@METR_Evals) November 5, 2026

4. We did not see convincing evidence of significant bottlenecks that would take effect within the next 12 months in algorithmic progress, training compute, experiment compute, or data. However, our view changed throughout the assessment, so we may be missing some bottlenecks.

— METR (@METR_Evals) November 5, 2026

What to do

Nothing changes in your Claude account today. A few habits help when tools move this fast:

1. Recheck your AI tools every few months. A model that fell short on a task in the spring may handle it now.
2. Pay monthly before committing to a year, since plans and models can shift quickly.
3. Keep your own notes on what works, like saved prompts and examples, so switching to a newer model is quick.
4. Read the report's limits section before sharing the "2x" figure. METR's own range runs from about 1.3x to 3x.