The ECI is computed from over 50 benchmarks on Epoch’s Benchmarking Hub, which combines Epoch’s internal evaluations with scores reported by model developers and benchmark creators. The index currently covers 168 models released since January 1, 2023. We restrict attention to models on the ECI frontier, meaning those that surpassed every earlier model on release. We limit analysis for this Data Insight to models released since March 14, 2023, because earlier models have especially sparse data and correspondingly noisy ECI estimates.

For each frontier model, we measure the number of days between its release and the release of the next frontier model. The current leader, GPT 5.5 Pro, has a lead measured only up to the latest available data.

GPT-4 led for 352 days, from its release on March 14, 2023 until Claude 3 Opus surpassed it on February 29, 2024. This is by far the longest lead in the data. The next-longest belongs to OpenAI’s o1, at 98 days, making GPT-4’s lead about 3.6 times as long. Among the 17 models that have led since GPT-4, none has held the top for a three-digit number of days, and the median lead has been about seven weeks.

• No pre-2023 coverage: Our public ECI does not score pre-2023 models, so we can only compare GPT-4 to models released after it. Models released before 2023 may also have led for extended periods, though we estimate this is unlikely, based on provisional high-uncertainty ECI scores.
• No significance threshold: We rank models by their point-estimate ECI and treat any increase as a change at the top. Some leads since GPT-4 end on small ECI differences that fall within confidence intervals. If we considered only statistically significant changes, several lead periods would lengthen moderately.
• ECI is periodically refit: As benchmarks and models are added, historical scores can shift slightly between updates.
• The current lead: The latest lead is measured only up to the most recent data, meaning it could extend beyond its current length.