Is AI speeding up AI research? METR's Anthropic report

On 5 November 2026, the independent evaluation group METR published its estimate that Anthropic is improving its AI models at about twice the rate it did between March 2024 and October 2025. METR guesses that about 80% of that speed-up comes from Anthropic using AI in its own research. It is the first public, outside assessment of a long-feared idea: AI that speeds up the making of better AI.

The report comes with large caveats. Anthropic redacted most of the evidence METR found most useful. METR says it would not be surprised if the true speed-up were as low as 1.3 times or as high as 3 times. This post explains what METR measured, what it could not see, and why the question matters.

Why does AI speeding up AI research matter?

If AI systems help build their own successors, progress could feed on itself. Each better model would speed up work on the next one. METR puts the worry plainly: this "could lead to a sudden emergence of extreme and risky capabilities", giving society less time to notice and respond.

Many AI developers treat this as a warning sign. Anthropic's Responsible Scaling Policy lists AI R&D acceleration as a key risk indicator. METR notes, though, that this assessment did not test against the specific thresholds in that policy.

What did METR actually do?

METR, short for Model Evaluation and Threat Research, is a nonprofit that studies the most advanced AI agents. It says it does not accept funding from AI companies. From 12 August to 29 October 2026, its staff, equal to 2.5 full-time people, ran what it calls a high-access evaluation at Anthropic. They asked three questions:

1. 1Has the growth rate of Anthropic's internal capabilities sped up?
2. 2How much of any speed-up is due to AI?
3. 3Are there bottlenecks that could slow growth in the near term?

Anthropic shared internal benchmark scores for its models, including unreleased ones, along with data on algorithmic progress and compute use. It also shared staff numbers, internal surveys of how much AI helps each team, lines of code and pull requests, and spending on AI agents. METR also ran about 15 interviews with key staff.

The access had limits. METR did not get to run unreleased models itself. Anthropic could approve or reject any request, and work was done on Anthropic devices. METR took no payment; Anthropic provided free model usage.

How do you measure research speed-up?

METR's main tool was a benchmark trend. Anthropic keeps an internal version of the Epoch Capabilities Index, which combines many benchmark scores into one number. METR fitted a straight line to the scores over time, allowed the slope to change at one point, and compared the slopes before and after.

The result: since October 2025, Anthropic's benchmark scores have been rising about 1.9 times faster than before. The 95% confidence interval runs from 1.5 to 2.8 times. In the chart, the trend went from about 15 points a year to about 29 points a year. METR's summary: Anthropic has made "roughly two years of historical progress over the last year". In other words, the past year brought about as much benchmark progress as the two years from October 2023 to October 2025, roughly Claude 2.1 to Claude Sonnet 4.5.

METR then checked whether the result held up when it changed its assumptions, such as which benchmarks to include and how to convert scores. The estimates ranged from 1.1 to 2.9 times. METR says it would be surprised by either no speed-up or one above 3 times.

If the statistics here are unfamiliar, our course on statistics for researchers covers confidence intervals and fitting trends.

How much of the speed-up is due to AI?

This is the harder question. Progress could speed up for other reasons, such as more computing power, more staff or better research ideas. METR's conclusions:

• Not more training compute: METR says it "received convincing evidence" that the speed-up was not caused by faster growth in training compute. The details were redacted.
• AI is the main cause now: METR guesses that a majority, "perhaps ≈80%", of the October 2026 speed-up is due to AI.
• Not always: the speed-up in late 2025 was likely influenced by factors other than AI.

Staff reports pointed the same way but are weak evidence. In informal Slack polls, some staff reported AI making them as much as 12 times more productive by October 2026. METR calls this "only very loose evidence". Interviews centred on about a 2 times speed-up. METR also warns that one person working faster is not the same as the whole company progressing faster.

Is AI doing research on its own? Not yet, says METR. AI "is not yet able to make highly significant and fully autonomous contributions to the most important parts of the model training process", but researchers described "signs of life".

What could not be checked?

This is the weakest part of the report, and METR says so itself. Anthropic had the right to redact any non-public information. Its stated aim was to remove details that could help competing AI developers.

• Most key evidence hidden: "Anthropic redacted a majority of the evidence we found most informative for these conclusions."
• Missing from the public report: detailed speed-up estimates by research area, the split between compute, data and algorithms, and forecasts for future models.
• Trust required: METR writes that "the public has to rely on METR's judgment much more than would be ideal."
• Scope cut: the original contract included checking how Anthropic splits compute between capabilities and safety work. Anthropic removed that part midway, before METR reached any conclusions, saying it wanted to stay focused on risk assessment for now.

METR also lists reasons the true speed-up could be lower. Non-AI factors might explain most of it. The speed-up might apply only to well-defined tasks that benchmarks measure. And Anthropic's earlier growth was unusually slow, which makes the jump look bigger. There are reasons it could be higher too. Progress might be curving upward, and benchmarks might understate real-world impact.

What does this mean for AI safety?

The most important line may be about bottlenecks. METR found no evidence of major limits on algorithms, compute or data that will take effect in the next 12 months. It concludes that "there is a significant risk that further capabilities progress could lead to even faster acceleration." It adds that its view "changed several times" during the work, so it is not confident.

Our view: this report is a useful first step in two ways. It gives a number, about 2 times, with honest error bars. It also shows that outside checks of AI labs are possible, if the lab agrees. But a check where the lab decides what the public sees cannot carry the full weight that safety decisions need. METR itself says staff-level access would allow a much deeper assessment.

For background, see our explainers on reasoning models, AI guardrails and whether AI can do research.

Research ideas for students

You cannot see inside a frontier lab, but public data lets you study the same questions:

• Trend breaks: take public benchmark results for open models over time and test for a change in slope, with confidence intervals. Does the answer depend on which benchmarks you include, as METR found?
• Measuring uplift: survey people in your lab about how much AI tools speed them up, then compare with an objective measure such as tasks finished. How far apart are they?
• Benchmark versus reality: find a skill where benchmark scores rose fast but real-world use did not change much, and explain why.
• Audit design: METR published how its access was set up. Design a better arrangement that protects company secrets while giving the public more evidence.

If you want to turn one of these into a project, our guide to starting research as a student explains how to find a mentor.