METR Finds Anthropic’s AI Benchmark Progress Sped Up 1.9×
In a November 5 assessment, METR estimated that Anthropic’s model benchmark scores improved 1.9 times as fast after October 2025. The underlying cause remains uncertain.
Explore the six-scene explainerKey takeaways
- METR estimated that Anthropic’s benchmark scores improved 1.9× as fast after October 2025 as they had before then.
- The estimated speedup has a 95% confidence interval of 1.5× to 2.8× on the benchmark.
- METR judged that AI assistance likely explains most of the recent acceleration, with substantial uncertainty and limited public evidence.
Sourced figures and status
| Period | Figure or status | Source |
|---|---|---|
| Internal assessment · 12 Aug–29 Oct 2026 | METR reviewed internal evidence and interviewed Anthropic staff. | METR assessment · 2026-11-05 |
| Benchmark trend · since Oct 2025 | 1.9× the earlier rate of score gains. The 95% confidence interval is 1.5× to 2.8×. | METR assessment · 2026-11-05 |
| Publication · 5 Nov 2026 | Anthropic redacted most of the evidence METR considered most informative. | METR assessment · 2026-11-05 |
What does 1.9× measure?
METR compared the rate at which scores improved on Anthropic’s adaptation of the Epoch Capabilities Index. The estimate means about 1.9 years of gains at the earlier pace took one year at the later pace. The analysis includes unreleased models. Sparse data and how the benchmark was constructed can affect the result. Source and context.
What did researchers learn about the cause?
Interviews and surveys pointed to AI assisting researchers. METR also examined training compute and other possible drivers. It judged that AI likely accounts for most of the recent speedup, but the precise contribution remains uncertain. The Associated Press reported that Anthropic co-founder Sam McCandlish agreed that AI was speeding up research and said the size of the effect was hard to pin down. Source and context.
Charts, maps and diagrams from the explainer





Frequently asked questions
What did METR assess?
It studied how much AI was speeding up AI research and development inside Anthropic. One finding was an estimated 1.9× increase in the rate of benchmark-score improvement.
Does the 1.9× estimate apply to all AI tasks?
The estimate concerns a particular benchmark for Anthropic models. METR says it is less certain how the trend applies to broader, less clearly defined tasks.
Sources and reporting scope
METR assessment. 2026-11-05. Primary report with the benchmark estimate, confidence interval, evidence on AI assistance and details of what Anthropic redacted.
Associated Press. 2026-11-05. Independent coverage of METR’s findings, how it obtained internal access and Anthropic’s response.
As of 7 November 2026. The 1.9× estimate compares rates of benchmark-score improvement on Anthropic’s version of the Epoch Capabilities Index (AECI). METR’s judgment about AI’s contribution is more uncertain, and Anthropic redacted much of the most informative underlying evidence.