Anthropic’s AI development may be moving twice as fast as it did a year earlier, according to a heavily redacted assessment by the nonprofit safety group METR. The group examined internal company data for about 11 weeks and concluded that Anthropic’s progress on certain tasks has accelerated, with AI itself driving much of the change.
METR’s report offers a rare look inside Anthropic’s research process. The group said it received “unprecedented access to internal data” and employee interviews, allowing its researchers to study whether the company’s internal capabilities were improving faster and how much of that progress came from AI.
AI may be helping Anthropic build better AI
METR estimates that Anthropic is improving its models on “well specified tasks” around twice as fast as it did between March 2024 and October 2025. Its estimated range for the speedup runs from 1.3 to 3x, with the nonprofit’s central conclusion pointing to progress at about 2x the earlier rate.
The group credits AI with around 80% of Anthropic’s faster progress. Anthropic has said its Claude chatbot is helping the company develop newer and more advanced versions of itself, a claim that fits METR’s main explanation: Anthropic is using AI to improve AI.
That process could create a compounding effect. If AI systems help researchers build better systems, and those systems then assist with the next generation, the pace of development could rise without a matching increase in human effort. METR identified that possibility as a key concern.
Ryan Greenblatt of METR said he was “seriously wrong” after expecting a much slower speedup this year. His comment captures the report’s central surprise: the progress measured inside Anthropic moved beyond the pace the researchers expected.
What the redactions hide
Anthropic removed most of the strongest evidence from the public report. METR wrote, “While we were able to freely share our highest-level conclusions about risk, Anthropic redacted a majority of the evidence we found most informative for these conclusions.” The redactions limited how the group could explain its more detailed estimates.
Sam McCandlish, an Anthropic co-founder and safety lead, said the company withheld material to avoid “handing out a roadmap for others to go faster.” McCandlish also said Anthropic plans to share “a report with our own analysis of the speed of AI progress within Anthropic.”
Even with the missing evidence, METR said the report revealed two internal models that were ahead of Anthropic’s top-ranked Claude Opus 5.5. One of them, called “Model 4,” cleared 50% on all 15 tested benchmarks. The report does not provide the names of the other internal model or the redacted evidence behind the comparison.
METR also warned that its findings could overstate the full picture. The acceleration may apply only to “well-specified tasks” that can be graded automatically or with a language model, rather than “messy” tasks that are harder to define. The group also said Anthropic may have understated its earlier progress, which would make the latest speedup look larger.
No clear bottleneck in Anthropic’s progress
METR said it could “confidently” rule out the growth of computing power used to train AI models as the reason for the acceleration. The report also found no “bottlenecks,” such as limits on computing power or data, holding back Anthropic’s progress.
That does not mean Anthropic’s systems can work independently across the entire research process. METR wrote, “We understand that AI is not yet able to make highly significant and fully autonomous contributions to the most important parts of the model training process.”
Still, Anthropic researchers described significant “signs of life” that could point toward those abilities arriving soon. The concern is not only what the models can do today, but how their assistance could change the pace of future model development.
Anthropic’s IPO filing defined a danger threshold as “dramatically accelerated progress attributable to AI.” Monday’s update said that point may be close, adding weight to METR’s warning about AI-assisted development.
A public call to slow down
The findings arrive after Anthropic CEO Dario Amodei called for slower AI development. In his Sept. 12 essay, “We Must Pace the Frontier,” Amodei urged the AI industry to “slow the pace at which we improve the capabilities of AI models” so safety measures can catch up.
In the same essay, Amodei supported external evaluators embedded within leading AI labs and specifically highlighted METR as an appropriate third-party evaluator. OpenAI CEO Sam Altman also supported independent evaluators with employee-like access.
METR began communicating with Anthropic in mid-August. One report drew on internal data collected from mid-August through late October, and the nonprofit submitted its final public report on November 3rd.
The report does not prove that Anthropic’s AI development will continue at twice its earlier pace. It does show that, after examining internal data, METR believes Anthropic is likely improving model capabilities around twice as fast on tasks that can be clearly defined and tested. The missing evidence leaves important questions unanswered, but the warning is clear: AI may already be helping accelerate the work of building better AI.
Based on
- An inside look at Anthropic’s AI speedup — therundownai.beehiiv.com
- 3 Takeaways From the AI Report on Anthropic Everyone Is Talking About – Business Insider — businessinsider.com
- Anthropic is rapidly accelerating AI development despite CEO’s call to slowdown, group says | The Independent — independent.co.uk
- An Anthropic risk assessment from AI safety organization METR is out. | The Verge — theverge.com
- Anthropic | The Verge — theverge.com
