METR Updates AI Capability Tracking Chart: Claude Opus 4.5’s Autonomous Task Performance Significantly Exceeds Expectations
The AI research non-profit METR recently updated its widely followed AI model capability tracking chart. The chart focuses on the ability of frontier models to autonomously complete long-horizon software engineering tasks, with the overall trend showing exponential growth. Anthropic’s latest model, Claude Opus 4.5, performed exceptionally in the evaluation, with its capabilities significantly exceeding the projected trend.
Overview of the METR Long-Task Benchmark
METR (Model Evaluation & Threat Research) is a non-profit research organization focused on evaluating the capabilities and potential risks of frontier AI models. Its landmark chart primarily measures the ‘time horizon’ of large language models on software engineering tasks, i.e., the length of a human task that a model can complete independently with a 50% success rate. Since its initial release in March 2025, the benchmark has been continuously updated to include data from the latest models, aiming to provide the industry with an objective reference for AI capability progress.
Claude Opus 4.5 Performance Highlights
Released by Anthropic on November 24, 2025, Claude Opus 4.5 is their most powerful model to date. In METR’s latest evaluation, the model’s 50% time horizon reached approximately 4 hours and 49 minutes (95% confidence interval: 1 hour 49 minutes to a much longer duration). This means Claude Opus 4.5 can autonomously complete many programming tasks that would require a human nearly 5 hours, a significant improvement over previous models. At the same time, its performance surpasses the chart’s original exponential trend forecast, drawing industry attention.
Exponential Growth Trend and Evaluation Complexity
The METR chart shows a steady exponential growth in the ability of AI models to autonomously complete long tasks. This trend is based on continuous testing data from multiple frontier models. However, MIT Technology Review points out that while the chart is a significant benchmark, the evaluation results for specific models come with large error bars, and the actual situation is quite complex. METR itself also emphasizes that the benchmark task configuration and evaluation methods are still being optimized, and interpretation requires consideration of multiple factors.
The Significance of Advances in AI Autonomy
Claude Opus 4.5’s excellent performance on the METR benchmark highlights the rapid progress of current frontier AI models in long-horizon reasoning, multi-step execution, and autonomous work. This update provides the latest reference for researchers and industry observers, helping to more accurately understand the development trajectory of AI capabilities. It also reflects the continuous evolution of evaluation standards to adapt to the rapid improvement of model abilities.
