On November 5, 2026, independent evaluator METR published its public assessment of AI R&D acceleration inside Anthropic. After high-access review from August 12 to October 29, METR’s subjective conclusion is that Anthropic is likely improving model capabilities on well-specified tasks at roughly twice the March 2024–October 2025 pace — with wide uncertainty (METR would not be surprised by ~1.3× or ~3×).

Coffee summary

  • FACT (METR): AECI (Anthropic’s Epoch Capabilities Index variant) growth since October 2025 is estimated at 1.9× the prior rate [95% CI 1.5–2.8×].
  • FACT: METR is confident the acceleration is not from faster training-compute scaling.
  • CLAIM / JUDGMENT (METR): A majority of current acceleration is likely from Anthropic’s use of AI for R&D; METR guesses perhaps ≈80% with substantial uncertainty.
  • FACT: Anthropic redacted a majority of the evidence METR found most informative; METR says the public must rely more on METR’s judgment than is ideal.
  • CONTEXT: Staff Slack-poll self-reports of personal uplift reached very high averages (METR cites ≈12× for the most capable internal model in October 2026) — METR treats these as loose evidence only.

What happened

METR staff (~2.5 FTE) reviewed internal benchmarks, training metrics, interviews, and surveys under a contract that gave Anthropic unilateral redaction rights over non-public information. The public report still shares high-level risk conclusions: apparent ≈2× acceleration on well-specified tasks, AI as a major contributor, and no clear bottlenecks identified that would block further acceleration over the next ~12 months — though METR repeatedly flags uncertainty and possible missed bottlenecks.

Robustness checks on AECI still generally show acceleration, but link-function choices swing estimates from about 1.1× to 2.9×. METR also warns AECI focuses on verifiable tasks and may not map cleanly to messy real-world impact.

Why it matters

If frontier labs are using their own models to speed model research, capability curves can bend without a matching jump in training compute — the scenario policymakers and preparedness teams worry about. METR’s access is unusually deep for a public write-up, but redactions mean outsiders cannot fully audit the strongest evidence. Read it as a carefully hedged risk signal, not a precise stopwatch.

Who should care

  • Policy and safety researchers tracking AI R&D automation risk indicators.
  • Investors and strategists interpreting “AI speeding AI” narratives with measured uncertainty.
  • Engineers comparing Anthropic’s public system-card acceleration notes with an outside evaluator’s view.

Limitations

Heavy redactions; no public sample access to unreleased models in this engagement; compute-allocation verification was dropped from scope mid-project. Self-reported uplift ≠ org-wide acceleration. Historical Anthropic AECI growth may have been below industry frontier, which could inflate apparent acceleration. METR discloses process and COI details in the post — read those before citing.

What to do next

  1. Read METR’s executive summary and redaction table before quoting a single multiple.
  2. Prefer ranges (1.3×–3× subjective; 1.5–2.8× AECI CI) over a lone “2×” headline.
  3. Pair with Anthropic’s own RSP / system-card language rather than treating METR as a regulator finding.
  4. Watch for follow-on assessments with fewer redactions or multi-lab comparisons.

AIImpish Take

METR is saying something important and incomplete at once: acceleration looks real, AI assistance is probably a big driver, and the public still cannot see the sharpest evidence. Useful as a risk update — weak as a scoreboard for marketing.

Related on AIImpish

Sources: METR, Independent Assessment of AI R&D Acceleration at Anthropic; AP News summary.