METR: Anthropic AI R&D likely ~2x faster than 2024–2025 baseline; majority attributed to AI use

Why it matters

【Key points】

• METR’s 2026-11-05 primary: high-access AI R&D acceleration assessment at Anthropic (fieldwork ~Aug 12–Oct 29, 2026, ~2.5 FTE, Anthropic unilateral redaction).
• Subjective: as of Oct 2026, Anthropic likely improves on well-specified tasks at ~2× the Mar 2024–Oct 2025 rate (plausible 1.3–3×)
• majority (~80% guess) of current acceleration attributed to AI use in R&D, not faster training-compute scaling.
• Public: internal AECI ~1.9× since Oct 2025 (95% CI 1.5–2.8×)—~two years of historical benchmark progress in one; METR confident compute-scale acceleration is not the driver.
• Interviews + “signs of life” for more autonomous work; no clear next-12-month bottlenecks (high uncertainty).
• Heavy redactions → public must lean on METR’s judgment.
• Companion same day: Anthropic Aug 2026 Risk Report R&D section review.

【Our read】

Background

AI-accelerating-AI-R&D is a core RSP / loss-of-control indicator. Not another leaderboard dump—a redaction-constrained high-access audit; strong claims with a clipped evidence trail are themselves a governance signal.

Peers

• vs Anthropic system-card ~1.7× AECI, METR’s ~1.9× includes newer unreleased internals and rejects a one-time Mythos Preview jump. vs Epoch ECI industry-wide, Anthropic-only + internals inflate the multiple
• industry-inclusive baselines can look ~1.4×—report lists under/overestimate channels.
• No comparable public high-access R&D-acceleration audit from OpenAI yet.

Impact

• reframes “self-acceleration?” for policy/media
• watch RSP triggers, peer audits, bottleneck reversals, autonomous contribution share, AECI↔real-world decoupling. paper

【Verifiable】

• Primary
• Companion Risk Report R&D review on metr.org same day.

Site: neither URL present; not prior Oct METR posts. max e/1089. Date-only → catch-up GO.