The Pulse
METR Estimates Anthropic’s Model Capabilities Are Advancing at About Twice Their Earlier Rate
METR estimates that Anthropic’s models are improving on well-specified tasks at about twice their earlier rate, while warning that the estimate is uncertain. Its assessment also describes faster growth in benchmark scores and evidence that
AI.info Team · Written with AI · fact-checked ·
Most METR staff who worked on the project estimate that, as of October 2026, Anthropic was likely improving model capabilities on well-specified tasks at around twice the rate it did between March 2024 and October 2025. The assessment, published November 5, draws on an eleven-week review of Anthropic’s internal AI research and development process. METR says the estimate carries high uncertainty: it would not be surprised if the true acceleration were closer to 1.3 times or three times the historical rate. The group is also less certain about how that acceleration applies to broader, less well-specified tasks.
AECI-measured benchmark scores accelerated
Since October 2025, benchmark scores for Anthropic’s models, measured using AECI—Anthropic’s version of the Epoch Capabilities Index—have grown at an estimated 1.9 times the pre-October 2025 rate. METR gives a 95% confidence interval of 1.5 to 2.8 times. The reported trend rose from 15 points per year before October 2025 to 29 points per year afterward.
METR says Anthropic appears to have made roughly two years of historical progress on benchmarks in the past year. Its estimate is higher than the 1.7-times acceleration reported in the Haiku 5.5 system card, primarily because METR included data from newer, unreleased models. Robustness analyses produced acceleration estimates between 1.1 and 2.9 times. METR says it would be surprised if the true evolution of benchmark scores showed no acceleration or acceleration above three times.
The report cautions that the AECI result could reflect overfitting to the benchmark distribution or an artifact of the index’s methodology. METR also says it is not confident that AECI can reliably estimate recent research and development acceleration: the data points lag research progress and are somewhat sparse. The benchmark finding informed the group’s broader assessment, alongside other evidence that it could not disclose.
METR attributes much of the current acceleration to AI
METR says it received convincing evidence that the acceleration was not caused by faster growth in training compute. It believes non-AI factors likely influenced the acceleration in late 2025, but estimates that a majority—perhaps about 80%—of the acceleration as of October 2026 is due to AI. The report gives that judgment with substantial uncertainty and says AI’s contribution has grown more important over time.
Interviews and surveys were consistent with AI use speeding up capability progress. METR’s overall impression was that researchers’ estimates covered a wide range but centered around a twofold acceleration. Anthropic also shared informal employee responses to Slack polls about speedup from the most recent model. The report describes average self-reported uplift as reaching approximately 12 times in October 2026, while cautioning that the poll responses are only loose evidence.
METR distinguishes individual speedup estimates from acceleration in overall progress. Sustaining a particular rate of growth in model capabilities generally requires sustained growth in inputs such as labor or labor productivity, the report says. It says research and development is currently accelerated primarily through AI assistance rather than fully autonomous contributions. Researchers described signs that fully autonomous contributions could occur soon, with maturity varying substantially by area.
No clear near-term bottlenecks identified
METR says it did not see evidence of significant bottlenecks in algorithmic progress, training compute, experiment compute or data that would take effect in the next 12 months. However, its understanding of those constraints changed as it received more information, and it is not confident in its current assessment. The report says the lack of observed bottlenecks creates a significant risk that further capability progress could lead to faster acceleration over the coming year.
Redactions limit public scrutiny
Anthropic had the right to redact non-public information from the assessment. The report says Anthropic redacted a majority of the evidence METR found most informative for its conclusions, describing the redactions as “aimed at removing confidential information that could accelerate other AI developers.” METR says the redactions make it harder for the public to judge the evidence and reach its own conclusions, leaving readers to rely more heavily on the group’s assessment than it would consider ideal.
METR staff, working 2.5 full-time equivalents, conducted the evaluation at Anthropic from August 12 to October 29, 2026. Anthropic provided access to internal information on pretraining, reinforcement learning, training compute, benchmark scores, speedup analyses, forecasts and staff interviews. METR says the access broadly surpassed its expectations, though Anthropic could approve or reject its requests. The group’s review considered three questions: whether Anthropic’s internal capabilities growth had accelerated, what share of any acceleration was attributable to AI, and whether near-term bottlenecks could slow progress.