Epoch AI

dashboard

public(free to browse)

eval-code

partial-open(several org repos MIT/Apache - SWE-bench docker registry, eci-public, MirrorCode-data)

benchmark-data

mixed(FrontierMath gated/private to prevent contamination

source

partial(public dashboard and MIT component repos, but no end-to-end harness published and the FrontierMath problems are unpublished)

Epoch publishes its results dashboard openly and releases some evaluation components under MIT and Apache licenses. But FrontierMath, its headline benchmark, is deliberately private, and no single OSI-licensed harness runs the whole pipeline, which puts it closer to source-available than to fully open source.

• https://api.github.com/orgs/epoch-research/repos recorded 2026-08-13

  33 public repos in the epoch-research org. Licensed eval components include SWE-bench and SWE-bench-c7d4d9d4 (MIT), eci-public (MIT), MirrorCode (MIT), benchmark-stitching (MIT), epochutils (MIT), ftm (MIT) and timelines-direct-approach (Apache-2.0); MirrorCode-data and many others carry no license at all. There is no single end-to-end harness repo among them.

• https://epoch.ai/benchmarks recorded 2026-09-29

  The hub describes itself as "Explore trends in AI capabilities across time, by benchmark, or by model." and its FAQ says "We publish task definitions for" several of the benchmarks it runs itself, including GPQA Diamond and Mock AIME 2024-2025.

• https://epoch.ai/frontiermath recorded 2026-08-11

  Epoch describes FrontierMath Tiers 1-4 as 'a benchmark of several hundred unpublished, highly challenging mathematics problems'. The problems are held back by design, so the benchmark behind the public dashboard cannot be run by a reader.

• https://epoch.ai/frontiermath recorded 2026-09-29

  Epoch describes FrontierMath Tiers 1-4 as "A benchmark of several hundred unpublished, highly challenging mathematics problems", so the headline benchmark behind the dashboard cannot be run by a reader.

• https://github.com/epoch-research recorded 2026-06-04

  public org repos incl SWE-bench docker registry, eci-public (MIT), MirrorCode-data; mostly MIT/Apache-2.0

Epoch's benchmark results, especially FrontierMath, are widely cited across frontier-model coverage, and the public dashboard spanning roughly 60 benchmarks is a recognized capability-trends reference. Epoch publishes no visitor, traffic or usage figure for the dashboard, so this reflects citation reach rather than a measured count.

• https://epoch.ai/benchmarks recorded 2026-06-04

  public database of leading-model benchmark results positioned as authoritative AI-capability tracker

• https://epoch.ai/benchmarks recorded 2026-08-13

  Free public dashboard letting a reader 'explore trends in AI capabilities across time, by benchmark, or by model', spanning roughly 60 benchmarks in three groups. No traffic, visitor or usage figure is published on it.

Epoch AI runs a wide set of hard, frontier benchmarks with trusted, standardized results, particularly in math, across every major lab. It is not a general agentic-eval harness, and its held-out design limits self-serve reproducibility, so it trails lm-evaluation-harness, the general reference.

• https://epoch.ai/benchmarks recorded 2026-06-04

  operated benchmarks list: FrontierMath, GPQA Diamond, SWE-bench Verified, SimpleQA Verified, the hardest tier of MATH (out of five), OTIS Mock AIME, Chess Puzzles.

• https://epoch.ai/benchmarks recorded 2026-09-29

  Dashboard stat tiles list 401 models tracked and 86 benchmarks tracked, "Featuring both Epoch-created and external evaluations, covering mathematics, coding, and more." The FAQ says "We use the" Inspect framework for evaluations and runs most models "16 times on GPQA Diamond".