METR Says Anthropic Is Improving Its Models About Twice as Fast as Before, Mostly Due to AI

METR, a nonprofit that evaluates frontier AI, said on November 5, 2026 that Anthropic is likely improving its models on well-specified tasks at about twice the rate it did between March 2024 and October 2025. METR guesses that a majority of that speedup, perhaps about 80%, comes from Anthropic using AI to do its own research.

The findings come from an 11-week engagement, from August 12 to October 29, 2026, in which METR staff reviewed internal Anthropic data on how fast its models improve. The question matters for AI safety research: if AI systems start to speed up the development of better AI, METR writes, extreme and risky capabilities could emerge suddenly. Anthropic itself lists AI R&D acceleration as a key risk indicator in its Responsible Scaling Policy, METR notes.

What did METR find about Anthropic's pace of progress?

METR's main measure shows benchmark scores of Anthropic's models growing 1.9 times faster since October 2025 than before, with a 95% confidence interval of 1.5 to 2.8 times. Put another way, Anthropic made about two years of its earlier progress in the past year, METR says.

The measure is AECI, Anthropic's version of the Epoch Capabilities Index, which folds many benchmark scores into one number per model. On that scale, the trend rose from about 15 points per year to about 29 points per year, with the break estimated at October 7, 2025. The data include three internal-only models; the newest, called Model 4 in the report, finished training in September and October 2026 and scored 178.6.

METR's overall judgment is a roughly 2x speedup, and it says it would not be surprised if the true figure were closer to 1.3x or 3x. Alternative analysis choices gave results from 1.1x to 2.9x, mostly depending on how benchmark difficulty is modeled. METR also flags a measurement problem: Model 4 scores above 50% on all 15 benchmarks it was tested on, and the lack of hard benchmarks makes the team "somewhat nervous" about measuring progress from here.

METR attributes most of the gap with the 1.7x figure in Anthropic's Claude Haiku 5.5 system card to the newer unreleased models it included. The lower 1.4x figure for all developers could mean Anthropic simply lagged the frontier in 2025, which would inflate the 1.9x estimate, or that rivals' earlier models were more tuned to benchmarks; METR does not pick between the two readings.

How much of the speedup comes from AI itself?

METR guesses that a majority, perhaps about 80%, of Anthropic's current speedup is due to AI, a judgment it gives with substantial uncertainty. It says it is confident the acceleration did not come from faster growth in training compute, and that non-AI factors likely played a large role in late 2025.

Three lines of evidence that METR could publish point the same way. Interviews by METR and Anthropic with researchers gave a wide range of answers centered on about a 2x speedup. Informal Slack polls of Anthropic staff put self-reported uplift at about 12x for Model 4 in October 2026, up from about 1.2x for Claude 3.7 Sonnet, though METR calls this "only very loose evidence". The researchers also described "signs of life" for AI making fully autonomous contributions to model training, which METR says do not yet happen at a significant level.

METR found no clear bottlenecks in algorithms, training compute, experiment compute or data that would slow progress in the next 12 months. Because of that, it sees "a significant risk" that further progress leads to even faster acceleration, while noting that its view of the bottlenecks changed several times during the work.

Why did Anthropic redact most of the evidence?

Anthropic had the right to redact any non-public information and removed a majority of the evidence METR found most informative, according to the report. Anthropic described the redactions as "aimed at removing confidential information that could accelerate other AI developers," METR writes.

Redacted material includes METR's detailed estimates for parts of the training pipeline, its breakdown of progress by compute, data and algorithms, and its forecasts for a future model. METR says the report therefore does not achieve transparency about evidence, and that the public must rely on its judgment "much more than would be ideal." Anthropic also canceled, midway through the project, a planned check of how it splits compute between capabilities and safety work, saying it wanted to keep the engagement focused on risk assessment.

The terms differ from the setup Anthropic announced on October 15, 2026, when it gave six METR staff employee-like access and said it could only require redactions in predefined categories such as customer data or model architecture details. METR says the new report comes from a separate engagement that began in mid-August, under which Anthropic held unilateral redaction rights over all non-public information. The same METR team did not request access to sample unreleased models, and METR worked on Anthropic devices with free tokens but took no payment. METR is also reviewing transcripts from the Claude Mythos 5 cyber test incident, according to the UK AI Security Institute.

What we don't know yet

• What the redacted evidence shows, including METR's forecast for a future Anthropic model and its detailed estimates by part of the training pipeline.
• Whether the acceleration holds on messier real-world tasks, since METR's main measure covers well-specified tasks that are easy to grade.
• Whether other labs see similar speedups; METR assessed only Anthropic.
• When, or whether, Anthropic will revisit the canceled check of its compute split between capabilities and safety, and what METR's employee-like access team will publish.