FollowingDay 3in the edition since 6 Nov 20265 sourcesupdated 7 Nov 2026

METR's Anthropic AI capabilities audit

The gist

Latest: METR: Anthropic AI Capabilities Growing at ~2x Historical Rate

An independent evaluator with internal access found AI capabilities at a leading lab are growing significantly faster than the historical baseline, with AI-assisted R&D identified as the primary driver. The extent of evidence redaction limits public ability to assess the underlying findings.

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

The full picture

METR published an independent assessment on November 5, 2026, finding that Anthropic's benchmark scores, measured via AECI, grew at an estimated 1.9x the pre-October 2025 rate, with an uncertainty range of 1.3x to 3x. METR attributes approximately 80% of that acceleration to Anthropic's own AI use in R&D workflows and rules out faster growth in training compute as the primary cause. Anthropic staff self-reported roughly 12x productivity gains from Model 4, and researchers described signs of AI autonomously conducting key training work. METR found no evidence of significant bottlenecks in algorithmic progress, compute, or data that would slow capabilities progress in the next 12 months.

Anthropologic held contractual rights to redact non-public information and exercised them over the majority of the evidence METR found most informative for its risk conclusions. Redacted material included evidence of speedup connected to specific parts of Anthropic's model training pipeline and related discussion of acceleration estimates. A claim about inference compute usage was also redacted. Anthropic additionally canceled a planned audit of how compute is split between capabilities and safety work. METR concluded that Anthropic's public August 2026 risk report understated how fast development was moving.

How it developed

6 November 2026

METR published an assessment on November 5, 2026 finding Anthropic's benchmark scores grew at roughly 1.9x the pre-October 2025 rate, attributing about 80% of that to Anthropic's own AI use in R&D.

Anthropic exercised contractual rights to redact most of the evidence METR found most informative, including pipeline speedup data, and canceled a planned compute audit. Ryan Greenblatt and Ajeya Cotra called it a major update toward superintelligence within 1 to 2 years. Miles Brundage argued Congress needs near-immediate enforcement timelines and separately that OpenAI and Anthropic lack a large or sustainable lead over other developers.

First citedMiles Brundage on X Alignment Forum Rohan Paul on X metr.org

First in the edition · 6 November 2026

5 November 2026in the 7 Nov edition

METR's blog post on the November 5 assessment added that Anthropic's redactions of key evidence rested on contractual rights Anthropic held over non-public information, characterizing them as a structural feature of the engagement rather than informal requests.

The assessment had found Anthropic's benchmark scores growing at roughly 1.9x the historical rate, with redacted material including evidence tied to specific pipeline stages, related acceleration estimates, and a claim expressed in compute rather than API-price terms.

First citedMiles Brundage on X The Neuron

Sources

Miles Brundage on X

Alignment Forum