Study: AI Nearly Doubles the Pace of Anthropic Model Development

The latest:

Anthropic’s artificial intelligence models are gaining capability nearly twice as fast as they did previously, with the technology itself taking on a larger role in building its successors, according to an assessment published Thursday by the nonprofit METR. The reviewer put the acceleration at 1.9 times the pace recorded between March 2024 and October 2025, while flagging wide uncertainty around that figure.

Details:

• The finding: METR, short for Model Evaluation and Threat Research, concluded that as of October 2026 Anthropic is likely improving its models’ capabilities on well-specified tasks at around twice the rate it managed between March 2024 and October 2025. The comparison is based on benchmark performance rather than commercial releases.
• The uncertainty: The nonprofit attached an explicit caveat to its own number, saying it would not be surprised if the true acceleration was closer to 1.5 or 2.5 times the historical rate. The 1.9 figure is presented as a central estimate within that range, not a settled measurement.
• The method: The assessment ran 11 weeks between August and October and drew on confidential research, interviews and performance data from models Anthropic has not released publicly. METR is an outside reviewer, not part of the company. It did not publish a breakdown of which unreleased models were examined.
• The limit: METR stressed that AI is not yet capable of making highly significant, fully autonomous contributions to the most important stages of model training. The acceleration it measured reflects AI assisting human researchers rather than replacing them at the decisive steps.
• Internal estimates: Anthropic researchers have reported sharp productivity gains from using AI to write code and run experiments, with informal employee polls pointing to improvements of roughly 12 times. METR cautioned that those self-reported figures are unreliable and treated them separately from its own benchmark-based estimate.
• The concept: The findings bear on recursive self-improvement, or RSI, the scenario in which AI systems build more powerful versions of themselves without human intervention. METR framed its results as potentially paving the way toward that threshold rather than evidence it has been crossed.
• The safety angle: The study amplifies recent warnings that capability advances could outpace the work of assessing their risks. METR’s role is evaluation and threat research, and the pace it measured is the core of the concern rather than any single capability.
• What is absent: The published assessment did not attach a timeline for when autonomous contributions to critical training stages might become possible, and no response from Anthropic to the findings was issued alongside the release.

Background:

METR is an independent nonprofit that evaluates frontier AI systems for dangerous capabilities and publishes assessments based on access arrangements with developers. Its work is cited in debates over whether external review can keep pace with commercial model development.

Between the lines:

The 1.9 figure and the 12 times employee estimate measure different things, and METR drew that line deliberately: benchmark capability gains are verifiable, self-reported productivity is not. The nonprofit’s own uncertainty band, 1.5 to 2.5, is wide enough that the headline number functions as a direction of travel rather than a precise rate, which is why the caveat about autonomous training contributions matters as much as the acceleration itself.

What’s next

Watch for Anthropic’s response to the assessment, METR’s next evaluation cycle and whether future reviews narrow the 1.5 to 2.5 uncertainty band or report AI contributions reaching the critical stages of model training.