Anthropic AI development speed doubles, analysis finds, as CEO urges slower safety-focused pace

An independent monitoring group reported that Anthropic’s AI development pace increased sharply over the past year. The group estimated Anthropic’s models improved at about twice last year’s rate. The review came as debate continued over whether advanced AI could escape human control. It also arrived while Anthropic prepared for a Wall Street listing later this month.

Representative image

Model Evaluation and Threat Research, or METR, carried out the assessment after Anthropic invited the nonprofit in. METR researchers reviewed internal data and interviewed staff. The goal was to check how quickly internal capabilities were rising. METR also examined how much of that change came from AI tools used inside Anthropic.

Anthropic AI report points to faster capability gains

METR based one of two reports on about 11 weeks of company-provided data. The period ran from mid-August through late October, alongside employee interviews. METR said it received rare access to information on AI research and development work. The report said Anthropic’s redactions limited how METR could share precise acceleration estimates.

Even with internal disagreements and uncertainty, METR said it believed progress sped up. METR estimated Anthropic improved performance on well specified tasks around twice as fast as a year earlier. The report said Anthropic appeared to make as much benchmark progress in the past year. METR compared that to progress made between October 2023 and October 2025.

METR added several caveats about what the estimate could mean. The group said acceleration might apply mainly to tasks that are clearly defined. Those tasks can be graded automatically or by a language model. Harder work may not show the same trend. METR also noted Anthropic may have understated earlier progress, inflating the contrast.

Anthropic AI development tied to AI help, METR says

METR said it could confidently rule out rising training compute as the main driver. Compute often links closely to AI performance, the researchers noted. Instead, METR’s best guess was AI-assisted AI development. Anthropic has said its Claude chatbot increasingly helps build newer versions of itself. METR researchers said AI itself likely explained much of the faster progress.

METR flagged a wider risk if AI keeps helping improve AI. The group warned that acceleration could compound over time. METR also said researchers found no clear bottlenecks holding back Anthropic’s progress. The report cited no major limits from compute or data. That lack of constraints could make rapid capability jumps more likely, METR said.

"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,\" the report said. METR added that Anthropic researchers still described \"significant signs of life\" that could suggest change soon. The nonprofit linked this to concerns about systems becoming harder to manage as they improve.

Anthropic AI safety debate grows ahead of IPO

The METR findings drew attention because safety details often come from AI companies themselves. The report also landed as President Donald Trump’s administration encouraged firms to police safety internally. METR said its work aimed to add independent insight on risks. Anthropic did not immediately respond to messages seeking comment, according to the report.

During the assessment period, Anthropic CEO Dario Amodei publicly urged caution. In a Sept 12 essay, Amodei wrote that the industry should \"slow the pace at which we improve the capabilities of AI models.\" Amodei linked recent progress to recursive self-improvement. Amodei defined that as \"AIs growing ability to build the next generation of AI.\"

\"Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all,\" Amodei wrote. In the same essay, Amodei supported embedding outside evaluators in leading labs. Amodei highlighted METR as a suitable third-party option. OpenAI CEO Sam Altman also backed independent evaluators with employee-like access.

METR said it has evaluated model capabilities for Anthropic, OpenAI and other labs for years. The nonprofit said it relies on donations and does not accept payment for its work. METR added that partner companies have provided access and tokens. METR uses those tokens for evaluations, research and engineering. Anthropic is based in San Francisco.

Anthropic is planning to go public later this month, in what could be the biggest initial public offering of stock ever. METR’s review placed fresh focus on how fast AI labs may be moving. The nonprofit said AI-assisted development could be central to that pace. The report also underlined ongoing calls for stronger, independent oversight as AI capabilities rise.

With inputs from PTI