GPT-4 vs rivals: how long models top the Epoch Index

OpenAI's GPT-4 held the top spot on the Epoch Capabilities Index for about a year, a run Epoch AI researcher Jaeho Lee says is far longer than any model since. Since Claude 3 Opus dethroned GPT-4 in February 2024, the index's leadership has flipped frequently: the lead has changed hands 17 times and the median stay at the top per model is about seven weeks.

What did the Epoch analysis find?

The Epoch analysis measures model performance with the Epoch Capabilities Index, and it shows GPT-4 as an outlier that led the index for about a year. The index is a composite measure of language model performance. Since Claude 3 Opus removed GPT-4 from the top in February 2024, leader turnover accelerated: 17 changes of leadership and a median tenure of about seven weeks per model.

The chart accompanying the analysis can be read two ways. First, it documents how long rival labs needed to match GPT-4, underlining GPT-4's unusually large early lead. Second, it shows more frequent but smaller capability jumps between successive leaders after the era that began with reasoning models like o1-preview in fall 2024.

How does GPT-4's reign compare to other leaders?

GPT-4's roughly year-long lead dwarfs subsequent runs: OpenAI's o1 is the second-longest, holding the top spot for just over three months, which the analysis notes is less than a third of GPT-4's reign. After GPT-4 lost the top position, no single model has approached that duration; most leaders last only a few weeks, producing the median figure of about seven weeks.

The piece frames GPT-4 as a genuine outlier at launch, while the post-GPT-4 period shows fiercer competition among labs. The transitions after February 2024 are faster in cadence but smaller in capability jumps compared with the step up GPT-4 represented and the transition into the reasoning-model era around fall 2024.

Why it matters

Shorter tenures at the top mean model improvements are happening more often and in smaller increments. That raises the bar for any lab hoping to claim a sustained lead: outlier wins like GPT-4's year-long run look harder to repeat. For customers and researchers the churn increases the pace of change they must track when choosing which model to build on or benchmark against.

What to watch

Track two concrete signals: whether any model surpasses the three-month mark that o1 reached, and whether a successor produces a capability jump comparable in scale to GPT-4. Epoch AI's count of 17 leadership changes since February 2024 gives a baseline for how fast the leaderboard can move; a longer single tenure or a large capability leap would indicate a shift back toward less frequent, larger advances.