LLM benchmark scores by model, since 2019

• Window: models released from October 2019 to Epoch AI's latest update, most of them in 2025 and 2026. Dates are UTC calendar dates.
• Grain: one row per published result, so a model can have several rows on one benchmark when the shots, harness, reasoning effort or source differ.
• Cadence: the table is rebuilt every day from Epoch AI's download. Epoch AI changed 36 of its benchmark tables between 29 September and 1 October 2026.
• New benchmarks: a benchmark Epoch AI adds to its Capabilities Index appears automatically, with one best score per model family, until its full table is added to the build.

• result_id: Epoch AI's record id for the result, or a hash of the row where Epoch AI publishes none.
• benchmark, benchmark_release_date: the benchmark as Epoch AI names it, and the date it was published.
• model, model_version: the model family and the exact version evaluated, usually the developer's API name plus Epoch AI's reasoning-setting suffix.
• organization, country, model_access, model_release_date: the developer, its country, how the model is released (API or open weights) and when.
• training_compute_flop: Epoch AI's estimate of training compute, in floating-point operations.
• score, score_unit, higher_is_better: the score as published, its unit and its direction. Units are fraction, percent, points out of 10, Elo rating, rating, speedup factor, Brier score, index points or US dollars.
• score_fraction: the score as a fraction from 0 to 1, for benchmarks scored as a proportion.
• stderr: the standard error the source reports, in the unit of score.
• result_date: the run, grading, addition or update date, where the source gives one.
• reported_by: Epoch AI for its own runs, Leaderboard, Model developer, Paper or report, or Not stated.
• source, source_url: the cited paper, leaderboard or site and its link, or Epoch AI's evaluation log for its own runs.
• evaluation_setup: the shots, harness, agent, reasoning effort or provider that tell several results for one model apart.
• notes, epoch_file: Epoch AI's note on the result and the file in its download the row comes from.

An empty cell means the source did not publish that value. Rows without a numeric score or a model name are left out, about 70 of the published rows on 1 October 2026. Results for older models often lack an organisation, access type or release date. Most results carry no standard error or result date. Epoch AI does not flag lab self-reports, so labs' own numbers appear under Paper or report and Model developer, next to academic papers.

• Ranking models on one benchmark
• Tracking how the best scores changed with model release dates
• Comparing Epoch AI's independent runs with numbers from leaderboards, papers and labs
• Relating benchmark scores to training compute and model access

Every row comes from the Epoch AI Benchmarking Hub at epoch.ai/benchmarks, read daily from its download. Epoch AI publishes its own evaluation runs under the Creative Commons Attribution 4.0 licence. Results Epoch AI compiled from other projects keep those projects' licences, so credit Epoch AI and the source linked in each row.