📅 November 5, 2026 • METR
METR — the nonprofit research group Model Evaluation and Threat Research — has published an independent assessment of how fast Anthropic's model capabilities are improving and what drives that speed. The report came out on November 5, 2026, and it is worth framing precisely: this is a third-party evaluation published by METR, not an Anthropic announcement. Over an 11-week period from August 12 to October 29, 2026, METR staff received unprecedented access to Anthropic internal data and staff, including metrics on pretraining, benchmark scores of internal models, forecasts of future performance, and interviews across the development pipeline.
The headline finding concerns pace. Per METR, benchmark growth on the AECI capability index has accelerated 1.9x since October 2025, with a 95% confidence interval of 1.5x to 2.8x. In historical terms, the group reads this as roughly two years of progress at the earlier rate compressed into a single year. The estimate carries explicit caveats: high uncertainty, and a trajectory measured on well-specified benchmark tasks, not every kind of real-world work.
The attribution question matters most for practitioners. METR estimates that AI itself accounts for perhaps ~80% of the current acceleration, a judgment the group marks as substantially uncertain. At the same time, METR confidently ruled out training-compute scaling as the cause — the speedup is not explained by faster growth of training compute. The picture is a lab whose research productivity rises because staff use AI heavily in day-to-day research work, not because models autonomously run the pipeline.
The publication drew immediate wire coverage. AP and Reuters both reported the findings on November 5, framing them as an outside check on whether AI development is now measurably speeding itself up. Both describe the access METR received as unusual for an external review of a frontier lab.
Impact: for AI labs and the teams that evaluate them, the report establishes that capability growth can be independently measured and attributed, and that a lab can grant an external nonprofit deep internal access with redactions disclosed openly — a template other labs may be pushed to follow. For teams building on Anthropic models, capability timelines may compress faster than roadmaps assume, which argues for frequent re-evaluation of what current models can do rather than annual planning cycles.
Sources:
https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/
https://apnews.com/article/anthropic-metr-ai-claude-acceleration-9cf8c590ffc1abdb489179bdccc69c23
https://www.reuters.com/technology/anthropics-models-improving-nearly-twice-fast-with-help-ai-study-finds-2026-11-05
Sources:
- https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/
- https://apnews.com/article/anthropic-metr-ai-claude-acceleration-9cf8c590ffc1abdb489179bdccc69c23
- https://www.reuters.com/technology/anthropics-models-improving-nearly-twice-fast-with-help-ai-study-finds-2026-11-05
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
