Analysis
METR estimates Anthropic's models now improve about twice as fast as in 2024-2025, with roughly 80% of the speedup tied to AI use
METR estimates Anthropic's models improve about twice as fast as in 2024-2025, with ~80% of the speedup tied to AI use; it flags big uncertainty.
On November 5, 2026, the nonprofit evaluator METR published an assessment of AI R&D acceleration at Anthropic. Its estimate is that as of October 2026, Anthropic is improving its models on well-specified tasks at about twice the rate it did between March 2024 and October 2025. METR's subjective view is that perhaps 80% of the current acceleration comes from AI use. METR itself calls that figure highly uncertain, says its team did not reach consensus, and says Anthropic redacted most of the evidence it found most informative.
The essentials
- The estimate: about 2x. The main benchmark-based figure is 1.9x with a 95% interval of 1.5 to 2.8x; METR would not be surprised by 1.3x or 3x.
- The 80%: a subjective judgment, "perhaps" 80% of the current acceleration, that METR calls highly uncertain. It does not mean 80% of progress, nor 80% confidence.
- What was measured: Anthropic's own AECI, a composite built from many benchmarks with automatically gradable answers.
- The caveats: limited access, redacted evidence, staff disagreement, and benchmarks that may overstate real-world acceleration.
What did METR actually measure?
Since August, a METR team worked with internal Anthropic information, over an assessment of about 11 weeks (AlphaSignal). The main quantitative evidence is AECI, Anthropic's version of Epoch's Capabilities Index, which fits a composite score to many benchmark results. AECI covers math, software engineering, science, multimodal and language tasks with checkable answers. METR notes this focus on clean, verifiable tasks may overstate real-world acceleration.
How solid is the "80% from AI" figure?
It is METR's subjective view and the team calls it highly uncertain. METR also says non-AI factors likely drove part of the acceleration in late 2025, and that the cumulative AI share since October 2025 would be lower than the current instantaneous one. Interviews and surveys are consistent with an AI-driven doubling, but staff-reported uplift, which reached about 12x in October 2026, is described as very loose evidence.
What is not established?
METR says it did not establish the exact AI contribution at each stage of training, because those estimates were redacted. It did not establish whether growth is accelerating further, that AECI reliably tracks real-world R&D progress, or that no bottlenecks will constrain progress over the next 12 months. It also says fully autonomous AI contributions to the most important training work are not yet large, though it saw "signs of life".
Summary table
| Point | What METR says as of November 5, 2026 | Source |
|---|---|---|
| Acceleration | About 2x vs March 2024 to October 2025, on well-specified tasks | METR |
| Benchmark estimate | 1.9x, 95% interval 1.5 to 2.8x | METR |
| Share due to AI | "Perhaps" 80%, highly uncertain | METR |
| Measure used | Anthropic's AECI, a benchmark composite | METR |
| Evidence | Much of the key evidence redacted by Anthropic | METR |
| Team view | No consensus among METR staff | METR |
Frequently asked questions
Does this mean Claude models are twice as capable?
No. It is a rate of improvement on well-specified, benchmarkable tasks, not a level of capability, and it may not carry over to real-world work.
Is the 80% a measured result?
No. METR calls it a subjective, highly uncertain judgment.
Method and limits
Sources as of November 5, 2026. We read METR's published assessment only through a summary of its page and cite The Information, Runtimewire and AlphaSignal for context. We did not review the underlying data, much of which Anthropic redacted. We tested nothing.
Sources
- METR: Independent assessment of AI R&D acceleration at Anthropic in October 2026 (verified 2026-11-05)
- The Information: Anthropic's R&D speed may have doubled due to AI, researchers find (verified 2026-11-05)
- Runtimewire: METR says Anthropic's AI capability growth rate has roughly doubled (verified 2026-11-05)
- AlphaSignal: METR finds Anthropic's AI-assisted R&D moving twice as fast (verified 2026-11-05)
Spotted an error in this article? Report it · Our corrections policy
