How long AI pauses would last based on ECI lag

Epoch AI's latest analysis quantifies the lag between leading AI labs using the Epoch Capability Index (ECI), a measure of model performance. The study compares each company's top model to OpenAI's frontier curve—tracking cumulative progress by release date—to estimate how long a pause would last before competitors catch up. For instance, Anthropic's Claude Fable 5 lags just 3 months behind, while Mistral and Microsoft trail by 20 and 26 months respectively. Some estimates (like Mistral's) may be inflated due to missing data, but the trend is clear: progress is unevenly distributed.

The analysis then flips the question: *How many companies must pause to achieve a given pause duration?* Assuming companies stop in order of their ECI scores, a 5-month pause requires only OpenAI and Anthropic to halt, as Moonshot AI (Kimi K3) would catch up in that time. Extending pauses becomes exponentially harder—9 months needs 7 companies to stop, and 18 months requires 11. This highlights the challenge of coordinated pauses, as even a modest slowdown demands broad industry cooperation.