METR Study Finds Anthropic AI Models Improving Twice as Fast
Model Evaluation and Threat Research found Anthropic AI models are improving nearly twice as fast as previous rates, suggesting AI is helping develop its own successors.
An independent study by the AI safety nonprofit Model Evaluation and Threat Research found that Anthropic AI models are improving nearly twice as fast as they did previously. The 11-week assessment, conducted between August and October 2026, determined that capabilities on well-specified tasks have increased 1.9 times faster since October 2025.
The findings suggest that AI is increasingly contributing to the development of its own successors, a trend that could lead toward recursive self-improvement. While the nonprofit noted that AI cannot yet make fully autonomous contributions to the most critical stages of model training, researchers at Anthropic reported significant productivity gains in coding and experimentation.
These results have amplified concerns among researchers that AI-driven acceleration could outpace risk assessment and make systems harder to align with human goals. Separately, Anthropic is preparing for a potentially massive initial public offering as soon as November 2026.
