Anthropic AI Acceleration Hits 1.9x on Benchmarks; METR Calls Anthropic Threshold Defense Overstated
Reported by Unite.AI, Crypto Briefing, Tech Times and 1 more outlets4 sources ↓
METR estimated that Anthropic’s benchmark scores improved 1.9 times faster after October 2025 than before, and said the company’s case that it had not crossed its 2x acceleration threshold was overstated.
Key points
- METR estimated benchmark scores rose about 29 AECI points per year after October 2025, versus 15 before
- METR put the acceleration estimate’s 95% confidence interval at 1.5x to 2.8x
- METR said crossing Anthropic’s threshold by July 15, 2026, was plausible, not established
- METR estimated about 80% of the acceleration came from AI use in Anthropic’s research
- Anthropic redacted most of the evidence METR considered most informative
Sources
- Unite.AIRead ↗METR Estimates Anthropic’s Model Capability Growth Has Nearly Doubled
- Crypto BriefingRead ↗METR suggests Anthropic’s AI development may have reached 3x acceleration threshold
- Tech TimesRead ↗Anthropic AI Acceleration Hits 1.9x on Benchmarks; METR Calls Anthropic Threshold Defense Overstated
- The DecoderRead ↗METR finds Anthropic may have crossed its own AI acceleration warning threshold back in July
Details
METR’s assessment compared Anthropic’s benchmark progress after October 2025 with its rate from March 2024 through October 2025. It estimated a 1.9x acceleration, with a 95% confidence interval of 1.5x to 2.8x; its robustness checks ranged from 1.1x to 2.9x under alternative modeling choices. The Decoder and Tech Times report annual AECI gains of about 15 points before the change and 29 afterward.
In a separate review, METR said Anthropic’s argument in its August 2026 risk report that acceleration was below its 2x threshold was overstated. METR said crossing the threshold by July 15 was plausible, while noting that most relevant researchers at METR and Anthropic leaned toward it not having been crossed. The review is not a finding that Anthropic breached its policy; the threshold also requires the doubling to be substantially attributable to AI research automation.
METR estimated that AI use in Anthropic’s research accounted for about 80% of the acceleration, but said this attribution was uncertain. It found the increase was not driven by faster growth in training compute. Informal employee reports of roughly 12-fold productivity gains were treated by METR as loose evidence that might overstate actual gains.
Anthropic had unilateral redaction rights over non-public information, and METR said most of the evidence it considered most informative was removed from the public report. The assessment ran from August 12 to October 29, 2026, with a team of 2.5 full-time staff and about 15 researcher interviews; METR said it took no payment from Anthropic.
Written from coverage as of
Images
- The Decoder
- Unite.AI
- Tech Times
- Crypto Briefing
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh