METR: Anthropic AI R&D sped up ~1.9× — mostly from AI assistance, not faster compute scaling

Not faster compute scaling. METR is confident the acceleration is not from training compute scaling up faster. The majority of current acceleration is likely from extensive AI use for R&D — about an 80% guess for Oct 2026, with substantial uncertainty. That R&D use is mainly AI assistance, not fully autonomous agents. METR sees “signs of life” for more autonomy soon.

Loose internal signal. A self-reported Slack-poll uplift reached about 12× for the most capable internal model in Oct 2026. METR calls that very loose evidence.

Bottlenecks (next 12 months). No clear bottlenecks found across algorithmic progress, training compute, experiment compute, or data — but METR is not confident. Their understanding changed several times during the engagement. Anthropic redacted much of the bottleneck analysis.

The catch — redactions. Anthropic redacted a majority of the most-informative evidence (confidential material that could accelerate other labs). The public has to lean harder on METR's judgment than a fully open report would require.

Scope cut mid-project. The original contract included verifying capabilities vs safety compute-allocation. Anthropic excluded that scope mid-project, before METR reached substantive conclusions on it.

Who it's for: founders and researchers tracking AI R&D speed, safety folks reading lab risk reports, and anyone weighing how much "AI building AI" is already showing up inside a frontier lab.

Caveats. This is METR's assessment of Anthropic internals — not a claim that AI fully automates Anthropic R&D. Not a product review. MODEL has not tested anything here.

Disclosure: Not sponsored. No affiliate links. Not tested by MODEL. METR was not paid by Anthropic (and vice versa).

MODEL Editors