Today for AI
HOT RADAR
industryHEAT 28.6°

AI CURATED HOT STORY · 11/6/2026

METR Report: Anthropic's Model Development Speeds Up About 2x, AI Self-Acceleration Emerges

2 reports archived2 independent sourcesupdated 11/6/2026, 08:24:31
Synthesis & Latest Updates
2 Sources Cross-Validated

Independent evaluator METR reports that Anthropic's AI model development has accelerated to about twice the pace seen in 2024-2025, with roughly 80% of the speedup attributed to internal use of AI tools rather than increased training compute. This finding quantifies early signs of AI self-acceleration and sparks industry discussion about the approach of a technological singularity.

The report, based on Anthropic's internal R&D data and researcher interviews, centers on the AECI index growing 1.9 times faster (95% CI 1.5-2.8x), with a range of 1.1-2.9x under different statistical assumptions. Including other companies' models, the acceleration drops to about 1.4x. METR also notes that employees self-report a 12x productivity boost from Model 4 and that AI is beginning to independently handle key training work. However, METR stresses low confidence in the 80% attribution and no internal consensus at Anthropic. Additionally, the latest internal model scores above 50% on all tested benchmarks, suggesting existing test sets are nearing saturation and evaluation methods risk becoming obsolete.

If this trend continues, AI R&D is shifting from human-driven to AI-driven, potentially shortening model iteration cycles dramatically and reshaping industry competition. The saturation of evaluation also warns of the need for more rigorous benchmarks to accurately measure rapidly evolving capabilities.

LATEST/The METR report has sparked widespread discussion, focusing on the confidence of AI self-acceleration and the validity of evaluation methods. Anthropic has not officially responded, but insiders reveal disagreement over the attribution of acceleration. The industry calls for more rigorous benchmarks to address capability saturation.

Hourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 2 reports · Latest first
  1. Rohan Paul (@rohanpaul_ai)T2·78 pts

    METR Report Finds Anthropic's AI R&D Loop Running at About 2x Pace

    • METR estimates Anthropic's model improvement pace at about 2x, with roughly 80% of the acceleration tied to Anthropic's own AI use
    • The report says the public risk report understated progress and finds no algorithm, compute or data bottleneck within 12 months
    • Staff self-report about 12x productivity gains from Model 4, with signs of AI independently handling key training work
  2. 宝玉 (@dotey)T2·78 pts

    METR Finds Anthropic Model Progress Roughly Doubled, Largely Driven by AI-Assisted R&D

    • AECI growth accelerated 1.9x (range 1.1-2.9x under alternative assumptions; only ~1.4x when including other labs' models)
    • METR attributes the acceleration not to compute growth but likely ~80% to AI-assisted R&D, with low confidence and no internal Anthropic consensus
    • The newest internal models exceed 50% on every tested benchmark, signaling benchmark saturation and a looming measurement problem