METR Assessment Finds Anthropic AI Improving Twice as Fast

Anthropic's AI models are improving far faster than they were in the previous period, and the company's own AI systems are doing much of the work behind that quickening. The METR assessment, published on November 5, puts the finding in plain terms: "As of October 2026, Anthropic is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025."

METR, a nonprofit that evaluates frontier AI systems, spent 11 weeks inside Anthropic for the review, working from August 12 to October 29. Its staff examined internal data on pretraining and reinforcement learning, training compute, benchmark scores for internal models, and analyses of how much AI use speeds up research, and interviewed staff. Anthropic gave the group access to data on unreleased models, according to METR, which said it took no payment for the work.

Benchmarks show the pace quickening

The central measure in the report is AECI, Anthropic's version of the Epoch Capabilities Index, which turns benchmark scores into a single trend. Since October 2025, that trend has grown at an estimated 1.9 times its earlier rate, with a 95 percent confidence interval running from 1.5 to 2.8 times. The trend rose from 15 points a year before October 2025 to 29 points a year after it, which METR reads as roughly two years of historical progress packed into the past year, about the distance from Claude 2.1 to Sonnet 4.5.

METR stresses the uncertainty in that headline figure. Its researchers said they would not be surprised if the true rate of improvement was closer to 1.3 times or 3 times the historical pace, and warned that benchmark scores can overstate or understate progress on less clearly defined tasks. The estimate also rests in part on evidence the public cannot inspect, a point the report returns to several times.

AI is doing much of the research

On why the pace changed, METR's subjective estimate is that a majority of the current acceleration comes from Anthropic's use of AI in its own research, perhaps around 80 percent. The group said other factors probably played a large role in late 2025, and that its attribution carries substantial uncertainty. It also said it was confident the quickening was not caused by faster growth in training compute.

Interviews with Anthropic researchers pointed the same way, with estimates centred on a doubling of the pace of capability gains. Informal polls of staff on Slack, in which employees voted on how much the latest model sped up their work, reached a self-reported twelvefold speedup in October 2026, though METR cautioned that the figures were unreliable and reflected little considered thought. The report also found no evidence that AI can yet make highly significant, fully autonomous contributions to the most important stages of training a model.

Much of the evidence stays hidden

The report also sets a limit on what readers can check for themselves. Anthropic held the right to redact any non-public information, and METR said the company removed a majority of the evidence it found most informative for its conclusions, describing the cuts as aimed at removing confidential information that could speed up rival developers. The group wrote that the public has to rely on its judgment far more than would be ideal, given the weight of the questions.

METR also reviewed the automated research section of Anthropic's August 2026 Risk Report and found the company's claim that it had not crossed its early-warning threshold for accelerating AI development rested on judgment calls that could be contested. The reviewers said it was plausible the threshold had already been crossed, and that the public version of the report understated the level of acceleration under way. The companion review is published alongside the main findings.

Why the finding matters

The assessment lands as Anthropic prepares to go public, potentially as soon as this month, in what Reuters described as possibly one of the biggest initial public offerings ever. The Associated Press also reported on the assessment. It also follows a September call by Anthropic chief executive Dario Amodei to slow the development of advanced AI. The company did not immediately respond to a request for comment, according to Reuters.

OpenAI, Anthropic's main rival, has released its own security-focused model under tight access controls, as covered in the report on GPT-6 Cyber.

The wider question the report raises is whether AI systems that improve their own development can keep quickening without the usual checks catching up. METR said it found no clear bottleneck in algorithmic progress, training compute, experiment compute or data that would prevent further gains over the coming year, and warned of a significant risk that further progress could lead to even faster acceleration. For now, the pattern METR describes is a loop: better models help build the next ones faster, while the evidence for how fast remains partly out of public view. The full report is published on METR's site, and the market context is covered in the report on the AI stock sell-off after the election.