METR Report Indicates Accelerated AI Development at Anthropic
A new report from METR, an independent research organization, suggests that Anthropic is experiencing a significant acceleration in artificial intelligence development. According to the findings, the company appears to be achieving approximately two years of progress within a single calendar year, a trend that researchers attribute largely to AI models assisting in the development of subsequent, more advanced models.
Performance of Unreleased “Model 4”
The METR assessment highlights the capabilities of an unreleased Anthropic model, referred to as “Model 4,” which was reportedly trained between September and October 2026. According to METR data, this model achieved a score of 178.6 on the AI Capability Index (AECI). This represents a jump nearly 90% as large as the entire performance improvement observed between the Opus 4.6 and Mythos 5 models.
Furthermore, METR reports that Model 4 scored above 50% on all 15 capability benchmarks tested. This performance has prompted discussions regarding the efficacy of current evaluation methods, with some observers noting that existing AI tests may be becoming too simple to accurately measure frontier progress.
Acceleration and Future Projections
Ryan Greenblatt of METR estimated that the current rate of AI research and development (R&D) acceleration has already reached levels previously predicted for March 2027 under various “AI 2027” forecasting scenarios. The report suggests there are currently no clear bottlenecks or obstacles preventing further acceleration in the field.
The data indicates that continued reinforcement learning (RL) on frontier mathematics and computer science continues to be effective in exceeding human-level performance. This progress appears to generalize significantly to other forms of planning and problem-solving.
Speculation on Model Lineage
While the METR report provides empirical data on model performance, some community discussions on platforms like Reddit have attempted to map the lineage of Anthropic’s released models to these unreleased versions. Unverified theories suggest that “Model 2” may have served as the teacher model for the released 5.1 series, while “Model 3” may have been the foundation for the 5.5 series, potentially through a distillation process.
Industry observers noted that scaled compute continues to provide a significant advantage, as it allows for a higher volume of agents to run simultaneously, thereby increasing the speed of model outputs. Despite the rapid pace of development, some analysts caution that while code is a vital metric for current capabilities, it may not be a complete measure of future model intelligence.