Written and maintained byCASRAI Editorial Board
Epoch AI is a data-first research nonprofit that publishes free datasets on the inputs and outputs of AI development: training compute, model characteristics, GPU clusters, chips, data centers and benchmark results. This guide is for researchers, analysts and policy staff who need to know which Epoch dataset answers which question, how to read its confidence labels and thresholds, and where its data stops. Everything below reflects Epoch AI’s own pages as read on 2026-11-07.
What Epoch AI is
Epoch AI describes itself as “a data-first research nonprofit investigating the future of artificial intelligence,” with a stated mission to help people understand what is happening in AI from a neutral perspective (epoch.ai/about). It began in 2021 as a volunteer group collecting AI data and became Epoch AI after a 2022 paper on training-compute trends. It is an independent, donor-supported organisation; its funding and team details sit on its own transparency and team pages.
Epoch’s published work may be used, distributed and reproduced with credit under CC BY 4.0. Its about page gives a preferred citation form and a BibTeX entry, so cite the version shown there rather than composing your own.
What its datasets cover
Epoch’s data hub (epoch.ai/data) gives each dataset its own page with a last-updated date and a download. Most downloads are ZIP files; some are CSV or XLSX. The hub does not state an update schedule, so check the date on each dataset page. The table maps research questions to datasets.
| If you need to know… | Use this Epoch resource | What the hub says about it |
|---|---|---|
| How training compute, parameters or dataset size have changed over time | AI models dataset | Over 3,700 machine learning models from 1950 to today |
| Where large training runs can physically happen | GPU clusters dataset | Over 500 GPU clusters and supercomputers (updated 6 November 2026 when read) |
| What accelerators exist and how they compare | Machine learning hardware dataset | Over 180 AI accelerators |
| How models score on hard tasks | AI benchmarks and capabilities hub | 97 benchmarks and 425 tracked models when read |
| Who owns or uses AI chips, and data-center build-out | Chip owners, chip users, chip sales and AI data center datasets | Listed as datasets; several are estimates (see limits below) |
| How people use AI | ChatGPT usage and AI-use-in-research datasets | The ChatGPT usage data comes from 5,000 US YouGov panelists, December 2022 to December 2025 |
The AI models dataset
This is the dataset most often cited. It tracks training compute, parameters, dataset size, training compute cost, training time and power draw, with an interactive graph (plot against publication date, colour by domain, organisation or country, filter by text or top N, and toggle regression and deep-learning-era views). Epoch defines three subsets: notable models (a state-of-the-art improvement on a recognised benchmark, over 1,000 citations, historical relevance or significant use), frontier models (the top five by training compute at release) and large-scale models (over 1023 FLOP). Downloads are available as CSV for notable, large-scale, frontier and all models, and the page includes a pandas snippet.
Benchmarks and the Epoch Capabilities Index
The benchmarks hub (epoch.ai/benchmarks) combines benchmarks Epoch runs itself, such as FrontierMath, GPQA Diamond, MATH Level 5, Mock AIME and SWE-bench Verified, with results gathered from external sources, each with its provenance cited. The Epoch Capabilities Index (ECI) combines scores from many benchmarks into a single general-capability scale; the hub notes an update on 4 November 2026 that moved it to a Bayesian model, so figures cited before that date may not match the current index. For internal runs Epoch uses the Inspect framework, each model’s API-default temperature and an empty system prompt, and shows error bars of one standard error across repeated runs. For the harness itself, see our guide to Inspect, the open-source evaluation harness.
How to use the data
- Open the data hub and pick the dataset whose subject matches your question, using the table above.
- On the dataset page, read the last-updated date and the methodology notes before downloading anything.
- For the AI models dataset, decide which subset you need: notable, frontier or large-scale. They answer different questions, and the thresholds differ.
- Use the interactive explorer to see the shape of the data first: choose the plotted metric, the colour grouping and the regression option.
- Download the CSV (or ZIP) for your subset. The AI models page offers a pandas snippet for loading it.
- Record the confidence label of each estimate you rely on (next section) and carry it into your analysis.
- Cite Epoch with the form on its about page and note the download date and the dataset’s last-updated date.
Strengths and limits
Strengths. Open licensing (CC BY 4.0), per-dataset download files, published task definitions and, for most internal benchmark runs, a log viewer of model interactions. The explicit definitions of notable, frontier and large-scale make subsets reproducible.
How to read the confidence labels. On the AI models dataset, estimates carry a label: Confident (within a factor of 3), Likely (within 10) and Speculative (within 30). A speculative compute figure can be off by a factor of 30, which matters for any regression or threshold test, so analyses should say which labels were included.
Limits stated by Epoch or visible on its pages.
- Coverage depends on public information. The explorer shows only models with estimates for the plotted metrics, and recent models may be missing; Epoch says major ones are typically added within two weeks, using a weekly automated search plus manual additions.
- Several datasets are estimates rather than counts: chip sales, data centers, owners, users and components.
- The ChatGPT usage and polling datasets are US samples.
- The AI-use-in-math-research dataset uses AI models to classify arXiv papers, so classification error is part of the data.
- On benchmarks, Epoch itself raises contamination (training data overlapping benchmark questions), sensitivity to small prompt or setting changes, and strict answer formatting that can push some models below the random baseline on GPQA Diamond. Only chat or instruct variants are included, not base models.
Alternatives and companions
Epoch measures inputs (compute, hardware) and capabilities in aggregate. It does not replace evaluation bodies that test specific models for dangerous capabilities. These CASRAI guides cover neighbouring resources, described as on those pages:
| Resource | What it is | Use it when |
|---|---|---|
| METR | An institutional profile of Model Evaluation and Threat Research: what it evaluates and how its evaluations work | You need evaluation of specific model behaviour rather than industry-wide compute data |
| Humanity’s Last Exam | A 2,500-question frontier benchmark by CAIS and Scale AI | You want a single named benchmark with human expert baselines |
| Inspect | An open-source evaluation harness, which Epoch uses for its internal benchmark runs | You want to run evaluations yourself |
For the wider context of why such measurement matters, start with what AI safety is, and for a separate annual overview see the Stanford AI Index guide.
Is Epoch AI free to use?
Yes. Epoch states its work is free to use, distribute and reproduce with credit under CC BY 4.0 (epoch.ai/about, read 2026-11-07).
Who runs Epoch AI?
It is an independent, donor-supported research nonprofit. Its funding and team are described on its own transparency and team pages, which should be checked directly for current details.
How do I cite Epoch AI data?
Use the citation and BibTeX entry given on the about page, and add the dataset name, its last-updated date and the date you downloaded it.
How many models are in the Epoch AI models dataset?
The data hub states over 3,700 machine learning models from 1950 to today, as read on 2026-11-07. The notable, frontier and large-scale subsets are far smaller.
What does Epoch’s “frontier model” mean?
The top five models by training compute at the time of release. “Large-scale” means over 1023 FLOP, and “notable” is defined by benchmark improvement, citations, historical relevance or significant use.
How reliable are Epoch’s compute estimates?
Each carries a confidence label: Confident within a factor of 3, Likely within 10, Speculative within 30.
Sources
- Epoch AI, About: https://epoch.ai/about (read 2026-11-07)
- Epoch AI, Data hub: https://epoch.ai/data (read 2026-11-07)
- Epoch AI, AI models dataset: https://epoch.ai/data/ai-models (read 2026-11-07)
- Epoch AI, AI benchmarks and capabilities hub: https://epoch.ai/benchmarks (read 2026-11-07)
Research-administration guidance, standards updates and independent tool reviews.
