Epoch AI Benchmarking Hub

What the hub is

The Epoch AI Benchmarking Hub is a database of benchmark results. It records how leading AI models perform on challenging tasks, and it brings together results from benchmarks that Epoch AI administers internally with data collected from external sources ^[1]. A benchmark, in this context, is a standardized set of tasks used to compare model behavior under similar conditions. The hub is therefore not one test but a collection: different benchmarks probe different skills, and the database puts their results in one place ^[1]. Because benchmarks differ in difficulty and format, a database that records them together can preserve the details of each test while still allowing cross-model comparison ^[1].

The Epoch Capabilities Index

The hub's Epoch Capabilities Index (ECI) combines scores from many different AI benchmarks into a single general capability scale ^[1]. Individual benchmarks usually focus on a particular kind of task. The ECI is designed to turn many such scores into one summary scale, making it easier to describe overall capability across benchmarks ^[1]. A single scale cannot preserve every detail of every test, but it gives a compact way to describe capability across a broad set of evaluations. The hub's homepage has tracked 425 models and 97 benchmarks, covering mathematics, coding, and other domains ^[1]. That breadth is the point: the collection spans more than a single evaluation area.

Internally run benchmarks

Epoch AI's internally administered evaluations began with GPQA Diamond and MATH Level 5 ^[1]. According to the hub, these were chosen because they were convenient to run, had not yet become saturated, and were frequently used by researchers ^[1]. A benchmark becomes saturated when scores pile up near the top and stop separating models; choosing tests that still leave room for differences makes the results more informative. Later, Epoch added Mock AIME 2024-2025, FrontierMath, and SWE-bench Verified to its internal evaluations ^[1]. These runs use the Inspect framework and published task definitions ^[1].

External results and scope

For benchmarks administered outside Epoch, the hub prioritizes those that best capture economically valuable real-world tasks, that are rigorously evaluated, and that are widely referred to by practitioners and researchers ^[1]. This is a selection principle rather than an attempt to include every benchmark. The hub also sets a clear scope for model versions: it evaluates only chat/instruct versions of models rather than base models, and it reports internal results only for models whose evaluations Epoch can run itself ^[1]. Those choices keep the internal comparisons tied to models Epoch can actually test, while the external data widens the picture ^[1].

Access and role

Epoch's data is free to use, distribute, and reproduce under the Creative Commons BY license ^[1]. The hub's role follows from its design: it gathers results from Epoch-run evaluations and external sources, organizes them across many benchmarks, and provides the ECI as a combined scale ^[1]. For a reader trying to follow progress in AI models, that makes the hub a reference point for seeing which benchmarks exist, which models have been tested, and how results look across mathematics, coding, and other domains ^[1].