• Products
  • Prices
  • Models
  • For Business
  • Materials
  • Company
Anthropic

METR: AI Has Nearly Doubled the Pace of Anthropic's Model Development

According to an , Anthropic's AI model capabilities have been improving 1.9 times faster on selected benchmarks since October 2025 compared with the previous period. Researchers estimate that approximately 80% of this acceleration can be attributed to the company's internal use of AI.For the study, METR gained access to Anthropic's internal data, including evaluation results from three unreleased models — Model 2, Model 3, and Model 4. The researchers also interviewed approximately 15 Anthropic employees.
METR: AI Has Nearly Doubled the Pace of Anthropic's Model Development
According to an , Anthropic's AI model capabilities have been improving 1.9 times faster on selected benchmarks since October 2025 compared with the previous period. Researchers estimate that approximately 80% of this acceleration can be attributed to the company's internal use of AI.For the study, METR gained access to Anthropic's internal data, including evaluation results from three unreleased models — Model 2, Model 3, and Model 4. The researchers also interviewed approximately 15 Anthropic employees.
METR: AI Has Nearly Doubled the Pace of Anthropic's Model Development

What the Data and Researchers Reveal

The nearly twofold acceleration was measured using the , based on a subset of tasks. The estimated acceleration ranged from 1.5 to 2.8 times within the confidence interval.
to Anthropic employees, Claude helps researchers write and review code, run experiments, analyze results, and test hypotheses more quickly.One researcher estimated that AI had reduced the time between developing an idea and running an experiment by approximately 80%. Another described the shift as follows: "The bottlenecks are now ideas and compute, not writing code."However, employees' assessments varied:
  • Two employees strongly agreed that development had accelerated by more than twofold.
  • Three employees generally agreed with this assessment.
  • Four employees provided mixed or inconclusive responses.
  • Two employees disagreed.
In separate surveys, employees estimated that working with Anthropic's most capable internal model had increased their individual productivity by an average of 12 times.METR considers these findings weak evidence of actual productivity gains, as they reflect subjective assessments rather than results from controlled experiments.
METR: AI Has Nearly Doubled the Pace of Anthropic's Model Development

Fully Autonomous AI Model Development Remains a Distant Goal

For now, Claude primarily accelerates individual stages of researchers' workflows. METR found no evidence of a substantial, fully autonomous AI contribution to the critical stages of training new models.However, researchers observed early signs that AI models are beginning to independently handle more complex tasks involved in model development.METR also identified no obvious constraints related to algorithms, computing resources, or data that would necessarily prevent further acceleration over the next 12 months.However, the researchers emphasized that this outlook remains highly uncertain.
METR: AI Has Nearly Doubled the Pace of Anthropic's Model Development

Contrasting Findings and Amodei's Warnings

The study comes after Anthropic CEO Dario Amodei in September for slowing AI development until safety measures could keep pace with advancing model capabilities.The findings also raise questions about transparency.Although Anthropic allowed METR to examine its internal data, the company withheld substantial portions of the underlying evidence supporting the study's conclusions.Anthropic explained that disclosing these details could provide competitors with a ready-made roadmap for accelerating their own AI development.Anthropic co-founder Sam McCandlish agreed with the study's central conclusion that AI is accelerating the company's research. However, he acknowledged that withholding parts of the underlying data makes independent verification more difficult.In a separate analysis of Anthropic's August report on the risks of automated AI research, METR also suggested that the company may have already crossed its own early-warning threshold for AI-driven development acceleration.Try Claude models on and explore their capabilities firsthand. 
METR: AI Has Nearly Doubled the Pace of Anthropic's :: BotHub