AI News
All stories
Reported

METR says AI speeds up Anthropic's R&D, AECI growing 1.9x faster since Oct 2025

METR's redacted risk report estimates that AECI scores for Anthropic models have grown 1.9x faster since October 2025, with a 95% interval of 1.5 to 2.8x.

Research · · Updated · 1 min read

Watch the short

Under a minute, read by an AI anchor. Made . The story has been updated since, and the text below is the current version.

YouTube loads only when you press play.

METR, after spending months with internal Anthropic data, has published a redacted risk report. It estimates that AECI scores for Anthropic models have grown 1.9x faster since October 2025, with a 95% interval of 1.5 to 2.8x.

METR's own caveats weaken that headline. The report says the estimate depends on analysis choices, and that the mapping from benchmark scores to real-world capability is unclear. Commentator @scaling01 makes a related point, arguing that a single AECI number tells readers little without per-benchmark disclosure.

Anthropic co-founder Sam McCandlish responded that METR had access to scores for unreleased models, internal projections and researcher interviews. He said the redactions were made for commercial sensitivity and to avoid giving others a roadmap. He also promised a separate Anthropic report on the speed of progress and on its standing under its Responsible Scaling Policy.

METR's Beth Barnes said the report takes no position on whether the redactions were justified. Daniel Kokotajlo objects on that question. He says a compute allocation assessment was dropped "to keep the focus on risk assessment," which he argues undermines verification of pacing. He also calls the redaction rationale inconsistent with Anthropic's pace.

Ryan Greenblatt, who also works at METR, reads the report as a big update toward artificial superintelligence within one to two years. He converts the roughly 2x overall acceleration into about 3.8x for AI software R&D, correcting his initial 4.5x figure, and places that rate around "March 2027" in the AI 2027 scenario. Jake Mendel now puts the chance of superintelligence within 18 months, absent a slowdown, above 50%. These are individual forecasts built on the same report, not findings of it. Peter Wildeford criticized Anthropic's stance on binding California commitments, and others pushed back on what they called the doomer reading of the results.

What changed

Type: Research · Applies to: Anthropic models

Why it matters

The report gives an outside, quantitative signal that a frontier lab may be using its own models to speed up research, which bears on how fast capabilities could move. Planners should weigh such estimates cautiously, since the authors themselves flag methodological limits.

How the story developed

  1. Adds the video's account of METR's engagement scope, the dropped compute check, and anonymous employee estimates of speedup, all attributed to the commentator.

  2. The full report adds the assessment period, METR's estimate that about 80% of the current acceleration is due to AI, the 1.3x to 3x plausible range, and the extent of Anthropic's redactions.

  3. This review adds METR's critique of the public report's wording, its own 1.3x to 2.8x range for leading indicators, and its view that the threshold may have been crossed.

Original source

Further reporting

Where we saw it

3 sources

Forward Future (Matthew Berman), Latent Space, METR

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works