METR estimates Anthropic's model progress has roughly doubled in pace, mostly from using its own AI

Anthropic redacted most of the evidence METR found most informative, and dropped a check on its safety compute halfway through.

By Think Facility · November 5, 2026 · Anthropic AI safety

On November 5, 2026, the evaluation group METR published what it found after eleven weeks with Anthropic's internal research data. Most of the METR staff on the project now think Anthropic is improving its models on well-specified tasks at around twice the rate it managed between March 2024 and October 2025.

METR's guess is that a majority of that speed-up, perhaps 80%, comes from Anthropic using its own AI to do the research. Anthropic redacted most of the evidence METR found most informative.

Who: METR, assessing Anthropic
Assessment: August 12 to October 29, 2026
Pace: around twice the March 2024 to October 2025 rate
Share due to AI: perhaps ≈80%
Benchmark trend: 1.9x since October 2025
Full Report: ~100 pages, delivered October 22

Where "twice" comes from

The cleanest number is from benchmarks. Anthropic keeps its own version of the Epoch Capabilities Index, called AECI, and its models' scores have grown 1.9 times faster since October 2025 than before (the 95% confidence interval runs from 1.5 to 2.8). METR puts that as about two years of the old progress in one year, roughly the distance from Claude 2.1 to Sonnet 4.5.

That's above the 1.7x in Anthropic's own Haiku 5.5 system card, mostly because METR also had scores for newer models Anthropic hasn't released. Robustness checks moved the figure anywhere from 1.1x to 2.9x. The team still disagrees internally, and it says the real rate could be nearer 1.3x or 3x.

We are confident that the acceleration was not due to an acceleration in the scaling of training compute.

From Independent Assessment of AI R&D Acceleration at Anthropic in October 2026

So it isn't just more chips. What's left could be hiring, better data or algorithmic finds, and METR thinks those non-AI causes likely mattered a lot in late 2025. Its view is that AI's share has grown since, which is where the 80% guess comes from.

What Anthropic's staff said

METR and Anthropic both interviewed researchers about how much AI had sped up their own area. The answers spread widely but centered on about 2x. Anthropic also runs Slack polls where staff vote with reactions on the speedup they get from the latest model, and those reached 12x in October 2026.

METR calls the polls very loose evidence, since nobody puts much effort into a Slack reaction. A 12x personal speedup also doesn't make the company 12 times faster; what METR watches is whether that number's growth rate changes.

For now it's mostly AI helping people. METR understands AI can't yet make highly significant, fully autonomous contributions to the most important parts of training, but researchers described "signs of life" that could change soon.

What Anthropic kept out

Anthropic had the right to redact any non-public information in the report, and METR had the right to say when it did. I read the redaction table first. Of seven kinds of evidence METR was shown, the public version describes the results of one in full and two in part.

Some of the key findings that most informed our conclusions about risk were redacted by Anthropic.

Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 - METR

Anthropic's stated reason is that the cuts were aimed at removing confidential information that could accelerate other AI developers. It sent redaction guidance on October 25 and tightened it between October 27 and 29, and the two agreed an extension of about five days. METR submitted the final report on November 3.

Anthropic also changed the job partway through. The contract first included checking how Anthropic splits compute between capabilities and safety. Anthropic dropped that before METR reached any conclusion on it, saying it wanted to keep the work focused on risk and would revisit compute verification later.

Why it matters outside the lab

METR spells out the stake at the top of the report. If AI starts to significantly speed up its own development, extreme and risky capabilities could appear suddenly. It found no clear bottleneck in algorithms, compute or data that would stop the speed-up compounding over the next 12 months, though it isn't confident in that.

For anyone outside the two companies, that's the practical problem. People can read METR's conclusions but can't check most of the evidence behind them, and METR says plainly that the report doesn't achieve transparency about evidence. Anthropic lists AI R&D acceleration as a key risk indicator in its Responsible Scaling Policy; METR didn't measure against those thresholds.

This isn't the arrangement Anthropic announced on October 15, when it said it would embed METR evaluators with employee-like access. This engagement began in mid-August, ran on Anthropic devices, took no payment from Anthropic, and didn't involve that kind of access.

METR published a second report the same day, a review of the automated R&D section of Anthropic's August 2026 Risk Report.