METR finds Anthropic model gains 1.9x faster, cites AI in R&D
The independent evaluation also names three unreleased Anthropic models, Model 2, Model 3 and Model 4, trained this year but not yet published.
METR published an independent evaluation report on Anthropic's AI R&D acceleration on November 5, finding that the company's rate of model performance improvement since October of last year is about 1.9 times faster than before, according to aitimes.com.
The organization analyzed Anthropic's internal data and model evaluation results over 11 weeks, from August 12 to October 29, and interviewed about 15 employees. It estimated that, as of October 2026, about 80% of the development acceleration came from using AI in R&D, and judged that faster growth in training compute was not the cause. Anthropic researchers use Claude for coding, experiment design and results analysis.
An internal survey found employees reported about 12 times higher productivity when using the latest internal model. Self-assessed gains were about 4x for Misos Preview, about 5x for Misos 5 and about 12x for the unreleased Model 4.
Three unreleased models confirmed
The report names Anthropic's internal models Misos Preview and Misos 5, and confirms three models the company finished training in July, August-September and September-October of this year but has not released publicly. They carry the temporary names Model 2, Model 3 and Model 4. It is not known which Claude product line they belong to or whether they will be released.
METR said including evaluations of the latest unreleased models is a main reason its 1.9x result is higher than the 1.7x Anthropic previously disclosed. It found early signs of AI contributing independently to model development, though most work is still directed and supervised by humans, and no clear compute, data or algorithmic constraints that would significantly slow the acceleration over the next 12 months.
METR said Anthropic may already have reached its own early-warning threshold for R&D acceleration as of July, though Anthropic judged it had not. Much of the report's key evidence was withheld at Anthropic's request, making independent verification difficult.
Quick answers
What did METR find about Anthropic's rate of improvement?
METR found Anthropic's model performance improvement since October of last year is about 1.9 times faster than before, above the 1.7x Anthropic previously disclosed.
Which unreleased Anthropic models does the report name?
Three models trained this year are labeled with the temporary names Model 2, Model 3 and Model 4. Their product line and release plans are unknown.
How much of the acceleration comes from AI?
METR estimated that as of October 2026, about 80% of the development acceleration came from using AI in R&D.