Understand this story
On November 5, 2026, METR published a blog post reviewing the "Risks from automated R&D" chapter of Anthropic's August 2026 risk report. METR believes that, as of July 15, the likelihood that internally deployed models would directly automate R&D in critical domains or trigger catastrophic risk was indeed low, but the report's claim that early-warning thresholds had not been crossed relied on contestable judgments not clearly explained in the public report; METR believes the models may already have crossed that threshold, and that the public report as a whole underestimated the degree of acceleration at the time. On the same day, METR also published an independent assessment stating that Anthropic's rate of model capability improvement since October 2025 has been roughly twice as fast as before, with benchmark scores growing 1.9 times faster based on AECI estimates, with a 95% confidence interval of 1.5 to 2.8 times, and 1.1 to 2.9 times under robustness checks, and judged that AI usage may explain most of the current acceleration.
AI-generated from public reports
Latest development · Report
METR report says Anthropic's internal model development has accelerated 1.9x since October 2025Developments
2 confirmed developments
METR reviews the automated R&D section of Anthropic's risk report
Source:METR Blog- Source:Testing Catalog
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 6
Testing Catalog
METR report says Anthropic's internal model development has accelerated 1.9x since October 2025
METR released a risk assessment report stating that Anthropic's internal AI model development has accelerated 1.9x since October 2025, attributing this to Anthropic's extensive use of AI in R&D. The report's chart classifies Opus 5.5 as a "non-frontier" model, while also listing internal models named Model 3 and Model 4, showing that it achieved substantial improvements in a very short period of time.
Testing CatalogRelated discussion
Testing Catalog joined this discussion
This post contributes to heat; it is not a formal report or confirmed fact.
Unite.AI
METR evaluation: Anthropic model capability growth rate nearly doubled
On November 5, 2026, METR released an independent evaluation concluding that Anthropic's rate of model capability improvement on well-defined tasks is roughly twice what it was previously, and judged that AI use may explain most of the current acceleration. Based on Anthropic's benchmark capability index AECI, METR estimates that since October 2025 the benchmark score has grown 1.9 times faster, with a 95% confidence interval of 1.5 to 2.8 times; under robustness checks, the acceleration estimate ranges from 1.1 to 2.9 times. The report states that growth in training compute is not the cause of the acceleration, but the relevant details were redacted, and Anthropic removed most of the evidence that METR considered most informative.
Nov 5
METR BlogFirst-party
METR reviews the "Automated R&D Risk" chapter of Anthropic's August 2026 risk report
After reviewing the "Automated R&D Risk" chapter of Anthropic's August 2026 risk report, METR considers that, as of July 15, the likelihood that internally deployed models would directly automate R&D in critical domains or trigger catastrophic risk was indeed low; however, the report's claim that the early-warning threshold had not been crossed relied on contestable judgments not clearly explained in the public report. METR believes the model may already have crossed that threshold, and that the public report as a whole underestimated the degree of acceleration at the time.