For this analysis we used the ECI data as accessed on 1 September 2026. The exact data used can be downloaded below.

Note for this analysis we did not extend the ECI to consider models that released before January 1st 2023, unlike our two prior works (1, 2), as we wanted to directly use the current ECI results.

The central trends are simple linear fits over each model that was SOTA on release for reasoning/non-reasoning models.

For simplicity we categorise every model released up to and including GPT-4.5 (27 February 2025) as non-reasoning, except a specific list:

• o1-preview
• o1-mini
• o1
• DeepSeek-R1
• Gemini 2.0 Flash Thinking (Jan 2025)
• o3-mini
• Claude 3.7 Sonnet

After GPT-4.5 we treat every model as reasoning.

The non-reasoning trend runs from GPT-4, which we judge to be the first genuinely frontier model in the standard ECI dataset, to GPT-4.5. The adj. R² for each trend is computed using the central ECI scores of its SOTA models: 13 reasoning, 10 non-reasoning.

The ribbons show a 90% prediction interval for the next model on each frontier. We construct them by first refitting the trend in each of the 500 ECI bootstrap samples, computing the frontier trend for each, and then incorporating the uncertainty for how models scatter around the line, so the predicted ECI at date \(t\) has distribution:

where \(D\) is the number of samples and \(a_d\), \(b_d\), \(s_d\) and \(\nu_d\) are the intercept, slope, residual standard deviation and degrees of freedom of the OLS fit in sample \(d\), with \(F_\nu\) the Student-\(t\) distribution function. Each ribbon runs between the 5th and 95th percentiles of this distribution.

The 90% interval on each rate comes from the same 500 refits, spanning the 5th to 95th percentiles of their slopes.