Breaking

METR Estimates Anthropic's Models Are Improving Twice as Fast, Attributes Most of the Gain to AI-Assisted Development; Anthropic Redacts Most of the Evidence

November 5, 2026 at 22:15 EST

Foundation Models
Research & Papers

Independent evaluator METR estimates that Anthropic's models are improving at roughly twice their previous rate on well-specified tasks, and that perhaps 80% of that acceleration comes from AI systems helping to build AI.

Third-party evaluation · METR on Anthropic

Anthropic's models reportedly improving at twice their prior pace, with most of the speed-up credited to AI helping build AI

An independent evaluation puts a number on how fast Anthropic's frontier models gain capability on well-defined tasks. Most of the reported acceleration is attributed to AI systems contributing to their own development, but much of the underlying evidence remains redacted.

2×

Reported rate of capability gain versus the previous rate, on well-specified tasks

~80%

Share of the acceleration attributed to AI systems helping build AI

Most

Of the key evidence is redacted, limiting outside checks of the figures

Rate of capability gain

Each block is one unit of prior pace. Column height is proportional to the multiplier.

1×

Previous rate

2×

Reported current rate

Where the acceleration is attributed

Each column is 100% of the reported acceleration, shown as 10 blocks.

~80%

AI helping build AI

~20%

Other factors

The remaining share is the complement of the reported 80% figure.

The reported feedback loop

AI assists engineering

→

Faster model development

→

More capable models

→

Better AI help

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.