METR Anthropic Report Finds AI R&D Moving About Twice as Fast

1.9 times. That is how much faster Anthropic's internal AI capabilities have grown since October 2025, according to the METR Anthropic report published on November 5. The Berkeley nonprofit spent about 11 weeks inside the company, and its central conclusion is blunt: AI is now doing a large share of the work that makes the next AI better.

The assessment arrives weeks before Anthropic's planned stock market debut and two months after CEO Dario Amodei publicly asked the industry to slow down. It also comes with a catch that METR itself puts front and center. Anthropic redacted most of the evidence the evaluators found most useful.

The METR Anthropic report in eight figures

• 1.9x: growth rate of Anthropic's internal benchmark scores since October 2025 compared with the March 2024 to October 2025 baseline.
• 1.5x to 2.8x: the 95% confidence interval around that estimate.
• About 2x: METR's overall judgment of the speedup on well specified tasks as of October 2026.
• 80%: METR's subjective estimate of how much of the October 2026 acceleration comes from AI itself.
• 11 weeks: the length of the engagement, from August 12 to October 29, 2026, against an original plan of about four.
• About 2.5: full time staff METR assigned to the work.
• About 15: Anthropic employees interviewed.
• 1.2x to 12x: the jump in staff self reported speedup from AI tools between January 2025 and October 2026, based on anonymous internal polls.

Two years of progress packed into one

The main number comes from AECI, Anthropic's version of the Epoch Capabilities Index, which rolls many benchmark results into one score. By METR's reading, Anthropic made as much benchmark progress in the past year as it did across the two years from October 2023 to October 2025.

METR says it would not be shocked if the true speedup were as low as 1.3x or as high as 3x. Across different analysis choices its estimates ranged from 1.1x to 2.9x. What it would find surprising is no acceleration at all, or anything above 3x.

The figure is also higher than the 1.7x Anthropic itself cited in the Haiku 5.5 system card. METR says the gap mostly comes from including newer models that have not been released.

What the 80% figure does and does not say

The most striking claim is about cause. METR says it is confident the speedup did not come from faster growth in training compute. Instead, it attributes roughly four fifths of the recent acceleration to AI systems helping Anthropic's researchers.

That help looks more like a strong assistant than an independent scientist. According to METR, current models cannot yet make highly significant, fully autonomous contributions to the most important parts of training. Researchers described "signs of life," the AP reported in its coverage of the assessment.

The 12x self reported figure deserves the least weight. METR calls the polling "very loose evidence," and interviews tended to land closer to a 2x speedup. People who use a tool every day are not always the best judges of how much time it saves.

This is the same dynamic Anthropic flagged in its own filings, which we covered in our look at the recursive self improvement warning in Anthropic's IPO paperwork.

No convincing bottleneck in the next 12 months

METR also looked for anything that might slow Anthropic down over the next 12 months: limits on algorithms, training compute, experiment compute or data. It did not find a convincing one. It also admits its view on bottlenecks changed several times during the work, so it holds that conclusion loosely.

The redactions behind the numbers

Here is where the report gets unusual. Anthropic had the right to redact non public information, and it used it heavily. Out went intermediate capability measures, algorithmic progress estimates, detailed bottleneck analysis, forecasts for future models and the stage by stage estimates of AI's contribution. Anthropic said the cuts were meant to remove confidential details that could speed up rival developers.

METR states in its published assessment that the public version does not achieve transparency about its evidence, and that readers have to lean heavily on METR's judgment. Redaction guidance arrived on October 25 and tightened over the following days, adding about five days to the schedule.

A planned check on how Anthropic splits compute between capabilities and safety work was dropped midway. Anthropic said it wanted to keep the focus on risk assessment and would revisit the topic later.

On independence, METR says it took no payment, though Anthropic supplied free tokens that METR used mostly for AI assisted analysis.

Why one evaluator's estimate matters now

Independent audits of frontier labs are becoming part of how governments plan to oversee AI. China is moving toward mandatory incident disclosure rules, and the US Justice Department just formed a task force on AI misuse. Labs are also building their own verification tools, such as Anthropic FlashProve.

METR's report shows both the promise and the limits of that approach. An outside group got deep access and produced a clear number. But the public cannot check most of the math behind it.

METR has said Anthropic should publish more of its benchmark data, and that harder benchmarks are needed to track progress that could soon outrun today's tests. Whether other labs invite the same scrutiny before their next releases is the question to follow on our AI news page.