On November 4, 2026, Epoch AI’s Emilia Alvarez and Alexander Barry published a methodology update to the Epoch Capabilities Index (ECI), its item-response-theory model that combines many benchmarks into one capability score. The model still estimates capability, benchmark difficulty and slope, but is now fit by sampling the full posterior with MCMC, producing 90 percent credible intervals instead of bootstrapped confidence intervals.
Other changes: a beta likelihood keeps predictions between 0 and 100 percent, each benchmark gets its own noise level, the raw scale is set from all benchmarks rather than WinoGrande alone, and slopes and noise get hierarchical log-normal priors. Claude 3.5 Sonnet stays anchored at 130 and GPT-5 at 150.
For the 176 models at or above ECI 125, estimates moved about 1 point on average (RMSE 1.1), every new estimate fell inside its previous interval, and intervals narrowed about 13 percent. GPT-6 Astra’s interval runs 166.2 to 172.7 and Claude Opus 5.5’s 165.8 to 172.0; Astra has a 72 percent probability of being more capable than Opus 5.5, 99 percent against Claude Fable 5.1 and 54 percent against GPT-6.1 Sol. The 76 models below 125 dropped 13.7 points on average, and GPT-4 fell from 126 to 122. Out of sample, 90 percent intervals covered 89.6 percent of held-out scores.
Why it matters: aggregate indices are increasingly how frontier progress is compared, and probabilistic head-to-head statements are more honest than point rankings when top models sit within a few points of each other.
What it does not show: uncertainty is relative to two anchor models treated as exact; weak models with few scores shift substantially, partly because old regularization was removed; coverage near 0 percent is only 80.5 percent and exact zeros cannot be predicted; and the steeper frontier trend slopes mostly reflect rescaling rather than faster progress.