On November 5, 2026, the nonprofit evaluator METR published what it calls an independent assessment of AI R&D acceleration at Anthropic. From August 12 to October 29, its staff had access to internal metrics on pretraining, reinforcement learning, compute, internal model benchmarks and staff interviews. Its headline conclusion is that Anthropic is "likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025." METR adds that it "would not be surprised if the true acceleration was closer to 1.3x or 3x the historical rate."
This is the outside evidence that was missing when this digest covered Anthropic's amended S-1 on November 3, in which the company said it might soon meet its own threshold for AI systems that speed up AI research.
What METR measured
- The number. Benchmark scores on AECI, Anthropic's version of the Epoch Capabilities Index, have grown at an estimated 1.9 times the pre-October 2025 rate, with a 95 percent confidence interval of 1.5 to 2.8. METR defines a 2x acceleration as two years of historical progress made in one year.
- Robustness. Alternative choices about which benchmarks and models to include gave estimates from 1.1x to 2.9x. METR's figure is above the 1.7x in the Haiku 5.5 system card, mainly because it includes newer unreleased models.
- The cause. METR guesses "a majority (perhaps ≈80%)" of the October 2026 acceleration is due to AI, and says it is confident the speedup did not come from faster growth in training compute. It also says non-AI factors likely influenced the speedup in late 2025.
- The ceiling. It saw no significant bottleneck in algorithms, compute or data due in the next 12 months, but says its view of bottlenecks "changed several times" and that it is not confident in that assessment.
Researcher interviews centred on about 2x. METR calls Slack poll estimates of personal speedup "only very loose evidence" and notes that individual speedup does not translate directly into an acceleration rate. Staff told METR that AI mostly helps as an assistant, and that fully autonomous contributions are not yet highly significant, although researchers described "signs of life".
The redaction problem
METR is unusually direct about its limits. Anthropic had the right to redact non-public information, and METR writes that Anthropic "redacted a majority of the evidence we found most informative for these conclusions." The company described the redactions as "aimed at removing confidential information that could accelerate other AI developers." METR's own conclusion is that "the public has to rely on METR's judgment much more than would be ideal."
What remains uncertain. AECI could reflect overfitting to the benchmark set, and METR says the data points are sparse and lag research. Benchmarks also cover verifiable tasks, so they may not reflect all the work that matters. The evaluation covers one lab, and it cannot be replicated from the public text.
Analysis: This is better than a company's self-report and weaker than a reproducible result. Treat the direction as credible and the magnitude as a wide range. For a smaller organisation the useful habit is the one METR models. When Jenya reports that Belgavi's Book Store's help-desk agent halved handling time on CASE tickets, the report should say what was measured, what was excluded and who could check it, because a confident number with its evidence removed is an opinion from someone well placed.
Sources and evidence
METR's high-access review estimates Anthropic's benchmark progress at 1.9x its earlier rate since October 2025 (95 percent interval 1.5 to 2.8, robustness range 1.1 to 2.9) and guesses about 80 percent of the current speedup is due to AI, not faster compute growth. Anthropic redacted a majority of the evidence METR found most informative, so outsiders must largely rely on METR's judgment.