Independent evaluators at METR reported on November 5, 2026, that Anthropic is likely improving model capabilities on well-specified tasks at about twice the rate seen from March 2024 through October 2025, with high uncertainty, and that benchmark scores on Anthropic’s AECI index have grown at an estimated 1.9× the prior rate since October 2025 (95% confidence interval 1.5×–2.8×). METR says a majority of the current acceleration is likely due to Anthropic’s use of AI in R&D, while stressing that Anthropic redacted much of the evidence METR found most informative. The assessment is central to TBR’s AI Safety and Ethics desk and should be read alongside company risk disclosures, not as a verified forecast of uncontrolled systems. Related Anthropic coverage includes Claude research-session statistics.
What METR concludes — with caveats
METR’s public post states three subjective risk conclusions: roughly 2× pace on well-specified tasks as of October 2026 versus the March 2024–October 2025 baseline; likely majority attribution of current acceleration to AI-assisted R&D; and significant risk that further progress could compound because clear bottlenecks were not observed for the next 12 months — while admitting the team may be missing bottlenecks. METR repeatedly flags uncertainty bands (including that the true multiple could be nearer 1.3× or 3×) and notes internal disagreements. AP News and The Verge summarized the same top-line findings for general audiences on November 5.
What the data shows
On AECI, METR estimates 1.9× acceleration since October 2025 with a 95% CI of 1.5×–2.8×, describing that as roughly two years of historical benchmark progress in one year. Robustness checks on link functions span about 1.1×–2.9×. METR says it can confidently rule out faster training-compute scaling as the cause. Staff self-reports of personal uplift reached very high averages in informal Slack polls (METR cites about 12× in October 2026) but warns those are loose evidence and do not equal organization-wide acceleration. Anthropic redacted detailed pipeline metrics and many of the strongest supporting exhibits.
Transparency limits
METR says Anthropic redacted a majority of the evidence the evaluators considered most informative, describing the company’s rationale as protecting information that could accelerate other developers. The public therefore must lean on METR’s judgment more than METR prefers. The engagement ran roughly August 12–October 29 inside Anthropic, with the final public report submitted November 3 and posted November 5. METR was not paid by Anthropic; Anthropic provided devices and token credits.
Why safety desks are watching
If AI systems meaningfully speed their own improvement, governance assumptions based on slower historical curves weaken. METR is not claiming recursive self-improvement is already autonomous end-to-end; it says assistance dominates today, with “signs of life” for more autonomous contributions. Policymakers and enterprise buyers should separate (1) benchmark acceleration, (2) real-world task impact, and (3) redacted internal evidence. TBR will update if Anthropic publishes its promised internal pace analysis or if METR releases less-redacted follow-ups.
How to use the numbers without overfitting
A 1.9× AECI acceleration with a wide confidence interval is evidence of faster benchmark growth, not a calendar for artificial general intelligence. METR itself warns that well-specified tasks may outpace messy real-world work and that Anthropic’s earlier AECI path lagged the wider industry frontier, which could inflate apparent acceleration. The right editorial use is to report METR’s range, attribution guess and redaction limits side by side — which this article does — and to update when either METR or Anthropic publishes less-redacted measurement.
Sources
Read next in Artificial Intelligence
Most read this week
- Abdul El-Sayed and the Muslim Brotherhood Claims: What’s Confirmed and What Isn’t
- Stock Futures Erase a 2.4% Overnight Drop as Markets Digest Democrats’ Sweep of Congress
- Spain Transport Minister Óscar Puente Quits 26 Days Before Election Over Crash That Killed 69
- GM Korea’s Sales Jump 26.6% on Chevrolet Exports as Hyundai Slips 6.3% and Korean Factory Growth Cools
- Boston Plans to Close 3 Schools in 2028 as Enrollment Drops as Much as 58%
Latest
- McGraw Hill Hooks K-12 Curriculum into Anthropic’s Claude for Teachers
- DeepMind’s Genie 4 Preview Builds Real-Time Worlds for Robot and Agent Training
- House Chairmen Ask Jensen Huang for Nov. 18 Briefing on Reported China Spy Link
- Google Backstops TeraWulf Debt for Anthropic’s $19B Kentucky Data Center Lease
- Nvidia Opens Four Cosmos 3 Nano Checkpoints for Driving and Robot Simulation