By Hamza Faisal
Anthropic's AI models are improving nearly twice as fast as they used to, and the acceleration is largely being driven by AI itself. That is the conclusion of an independent assessment published on November 5 by METR, the nonprofit group that evaluates AI systems, after 11 weeks inside Anthropic examining internal data.
METR's central finding is striking: since October 2025, Anthropic's benchmark scores have been growing at 1.9 times their earlier rate, with a 95 percent confidence interval of 1.5 to 2.8 times. In plain terms, the lab made roughly two years of historical progress in a single year. METR's broader judgment is that capabilities are improving at around twice the March 2024–October 2025 rate, though the true figure could be 1.3 to 3 times the old rate.
Two years of progress in one
The assessment ran from August 12 to October 29, 2026, with METR staff reviewing internal data on pretraining, reinforcement learning and benchmark scores, and interviewing about 15 staff members. METR says it took no payment from Anthropic, which provided laptops and free tokens for the work.
The headline number comes from AECI, Anthropic's version of the Epoch Capabilities Index, which combines scores across many benchmarks into a single measure. METR estimates the index rose from about 15 points a year before October 2025 to 29 points a year after it. That tops the 1.7 times in the Claude Haiku 5.5 system card, mainly because METR also counted three newer unreleased models. One of those, Model 4, was the first to score above 50 percent on every benchmark tracked.
METR says it is confident the speedup was not caused by faster growth in training compute. Instead, it estimates AI itself — models helping with research, coding and experiments — accounts for perhaps 80 percent of the acceleration, though with substantial uncertainty.
The evidence readers cannot see
The report comes with a major caveat. Anthropic had the right to redact any non-public information, and METR says the company redacted a majority of the evidence it found most informative, including its most detailed analysis of AI attribution and forecasts for future models. Anthropic said the redactions removed confidential information that could accelerate other AI developers. Outside researchers cannot independently check conclusions that rest on unreleased models.
METR also published a companion review of the automated-research section of Anthropic's August 2026 Risk Report. It agreed the risk from models as of July 15 was low, but said Anthropic's claim not to have crossed its “2x Acceleration Threshold” — an early warning for when AI begins accelerating its own development — rests on contestable judgment calls. METR said it was plausible the threshold had already been crossed, and that the public report understates the acceleration occurring.
Why the finding matters
The assessment is some of the clearest evidence yet that AI is accelerating its own development, a dynamic researchers call recursive self-improvement: systems helping to build more powerful versions of themselves. Reuters noted it amplifies warnings that such advances could outpace efforts to assess the risks.
METR found no convincing evidence that algorithmic progress, training compute or data would impose significant bottlenecks over a 12-month horizon, and warned that without clear constraints, capability gains could compound into even faster acceleration. The group cautioned, though, that AECI measures well-specified tasks, so benchmark gains may not mean equivalent real-world effects.
For now, the numbers point one way: the pace of AI progress at one of the world's leading labs has roughly doubled, and the tools doing much of the work are the models themselves.
