Tutorials & Tips

[Evergreen] How to read METR’s “~2× faster”: progress speed on well-specified tasks only

[Evergreen] Bottom line: METR’s “about twice as fast” means Anthropic’s rate of capability progress on well-specified (clearly graded) tasks as of October 2026 is roughly double its March 2024–October 2025 historical pace—not that all AI abilities doubled, and not that models finish quizzes in half the time. The public estimate is highly uncertain (~1.3×–3×); the AECI trend is about 1.9× [1.5–2.8]. Much supporting evidence in the public report is redacted.

Q: What are well-specified tasks? Clear goals and easy automatic or semi-automatic scoring. METR says extrapolating this acceleration to messier tasks is more uncertain.

Q: How can ~2×, 1.3×–3×, and ~1.9× coexist? About 2× is METR’s subjective synthesis for progress speed on well-specified tasks as of October 2026; 1.3×–3× is the range they would not be surprised by. AECI ~1.9× since October 2025 is a key empirical trend, not the only input.

Q: What does “acceleration” measure? A ratio of growth rates (2× ≈ two years of historical progress in one year)—not staff self-reported personal speedup, and not quiz completion time.

Q: Where does the speedup come from, and how much can the redacted public report support? METR is confident it is not from faster training-compute scaling. For current (≈Oct 2026) acceleration they guess a majority (perhaps ≈80%) is due to AI used in R&D, with substantial uncertainty; late-2025 acceleration likely mixed in non-AI factors. Redactions mean the public can retain the high-level judgment but cannot independently re-check evidence at the same strength.

Q: How is this different from other METR pieces on this site? Full-assessment short read https://new.nat6.net/news/custom-727783c45e414d2e/ ; Risk Report R&D chapter review https://new.nat6.net/news/k53ssxx8ie3n6fzyqmky4izv7/ ; warning-threshold reporting https://new.nat6.net/news/vpopkx98s6mpp9coxhi7hhvh2/ . This page only teaches how to read “~2× faster.”

How ordinary readers can use this Keep these limiters when reading headlines: well-specified tasks; progress speed (not quiz time); separate ~2× / 1.3–3× / AECI ~1.9×; heavy redaction in the public report. Do not extrapolate to “all AI abilities doubled,” “Claude overnight 2×,” or “full self-acceleration”—messier tasks are outside the same confidence.

Source: https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/

Quick Facts

Published
Source
METR
Category
Tutorials & Tips
Topic
Anthropic / ClaudeSafety & AlignmentIndustry News
Why it matters

SEO Brief 1 常青问答:教读者读懂 METR「约快一倍」的适用范围、三组数字与涂黑边界;区别于站内已发短讯。

Topics

Related entities

How to Cite

When citing this article, keep the title, date, and permalink, and credit Orange News as the source.

Orange News: [Evergreen] How to read METR’s “~2× faster”: progress speed on well-specified tasks only. November 7, 2026. https://new.nat6.net/en/news/custom-35f9e3fcc1784656/

View all →