Anthropic is developing its artificial intelligence systems about twice as quickly as it was a year ago, according to an analysis released Thursday by an independent monitoring group. The study found that AI itself is playing a larger role in improving newer models. METR researchers said they received unprecedented access to internal Anthropic data for nearly three months. They interviewed employees and examined data about unreleased models.
Anthropic is rapidly accelerating AI development amid slowdown calls, outside research group says
METR estimated Anthropic is improving capabilities for well specified tasks around twice as fast as a year earlier, AP News reported. Benchmark scores suggested about two years of historical progress over the last year. Researchers said they are confident the acceleration was not simply due to scaling computing power. That mattered because more computing has historically driven AI progress.
Anthropic invited researchers from Berkeley, California based nonprofit Model Evaluation and Threat Research to review internal data and interview staff. The timing came as Anthropic prepares for its Wall Street debut. The session also overlapped with public calls from Chief Executive Dario Amodei to slow AI development. In a September 12 essay, he urged the industry to slow the pace of capability improvement.
3 takeaways from the big AI report on Anthropic everyone is talking about
The report offered some of the clearest evidence yet that AI is accelerating its own development, Reuters reported. The findings potentially point toward recursive self-improvement, where systems build more powerful versions without human intervention. METR warned that fully autonomous contributions may play a larger role soon. It also found no clear signs of diminishing returns or observable bottlenecks preventing even faster acceleration.
Co-founder Sam McCandlish said many of METR conclusions match his own, AP News reported. He wrote on X that AI is speeding research, but exact magnitude is hard to say. For now, much acceleration appears to come from AI assisting human researchers. Claude, Anthropic chatbot, increasingly helps the company develop newer and more advanced versions, the company has said.
Anthropic recently finished 3 new models it has yet to release
Over the past six months, Anthropic finished training three models it has yet to release, Business Insider reported. The models are codenamed Model 2, Model 3 and Model 4. They completed training in July, August to September, and September to October respectively. METR also identified an unreleased model dating back to 2022. Those internal models help drive faster capability gains.
Anthropic workers are using better AI models than the public, the report said. Keeping advanced models private gives research and product teams a competitive advantage. Companies may also keep models private longer because of safety concerns. The disclosure reflects a central tension in selling advanced AI while retaining internal advantages.
Anthropic staff say their productivity is skyrocketing
METR wrote that employees using the most capable internal model self reported being about 12 times as productive in October. That was around double what staff reported a few months earlier while using Mythos 5. The evidence came from informal Slack polls and interviews with Anthropic employees. METR cautioned that Slack polls are weak evidence on their own. The figures are self reported and should not be treated as verified output measures.
Anthropic turned over internal Slack poll data showing major productivity gains, Business Insider reported. METR interviewed employees and surveyed Slack channels to build the report. The researchers received unprecedented access to internal data for nearly three months. That access, and its limits, shaped both conclusions and redactions in the public version.
Redactions limit what the public report can show
METR said Anthropic redacted a majority of the evidence it found most informative for its conclusions, Business Insider reported. The researchers said they could freely share highest level conclusions about risk. Anthropic was permitted to redact portions under its agreement with METR. The public report therefore offers a rare glimpse, but omits much supporting detail. Readers cannot independently verify several underlying claims from the released evidence.
METR has evaluated models for Anthropic, OpenAI and other labs for years, AP News reported. The nonprofit is funded through donations and does not accept compensation for its work. Partner companies have provided access and tokens for evaluations, research and engineering. The group is perhaps best known for its report on OpenAI Hugging Face agent breach.
Questions and answers
Did Anthropic cross its own AI R and D threshold? Not disclosed. No specific threshold, amount or crossing date was stated in the verified reports reviewed here. The brief also cautions against asserting a crossing without independent verification from primary METR or Anthropic documents. Treat any such claim as unverified unless primary documents state it directly.
How much of the acceleration does METR attribute to AI? METR said the majority of acceleration as of October is due to AI, and its contribution is growing. It did not provide a verified percentage in the reports reviewed here. Researchers said informal Slack polls and employee interviews supported sizable productivity gains. Those sources are useful context, but they are not hard output measurements.
Is this evidence that recursive self-improvement has started? No, not as a verified conclusion. The findings offer evidence that AI is accelerating its own development, Reuters reported. That potentially paves the way for recursive self-improvement in the future. METR said much acceleration still comes from AI assisting humans, so treat the question as open.
What did Anthropic redact from the METR report? Anthropic redacted a majority of the evidence METR found most informative, Business Insider reported. METR could still share highest level conclusions about risk. The exact content of redactions was not disclosed in the reports reviewed here. Specific redacted evidence therefore cannot be cited as verified fact.
What did METR conclude about Anthropic on November 5, 2026? METR concluded models are improving roughly twice as fast as a year earlier. It also found three unreleased models finished in recent months. Staff self reported about 12 times productivity with the most capable internal model. METR found no clear signs of diminishing returns or bottlenecks to even faster acceleration.
Where can I read the primary documents? Not disclosed. No direct METR report URL was provided in the verified sources used for this article. Readers should seek the primary METR and Anthropic documents directly before relying on secondary summaries. Redactions and unverified claims make primary source checks especially important here.
Anthropic plans to go public later this month, AP News reported. METR findings will likely remain central to debate over AI pace, safety and self improvement.




