Epoch AI
Epoch AI · Writing · Business
Epoch AI is a nonprofit research institute that tracks AI progress through the FrontierMath benchmark, the Epoch Capabilities Index, and open databases of AI models, chips and data centers.
About Epoch AI
What Is Epoch AI
Epoch AI is a research organization, not a software product. It collects hard numbers on how AI is built and how fast it improves, then publishes the datasets and the analysis for free. The work sits behind a lot of the charts you see in government reports and news coverage about AI scaling.
The group started in 2021 as volunteers cataloging machine learning models. A 2022 paper on training compute trends gave the project enough traction to become a proper institute, and it has run as an independent nonprofit since then, funded by donors and by commissioned research.
What it actually gives you is evidence. If you want to know how much compute went into a model, how a new release scores on a hard math test, or how much power the largest AI data centers draw, this is one of the few places with a sourced answer.
The limits matter too. Epoch AI is small, so coverage follows what its team can verify, and some datasets update in bursts rather than daily. FrontierMath keeps most of its problems private, which protects the test but also means you can't rerun the full set yourself.
Getting Started
- Open epoch.ai and choose the area you need: AI models, capabilities and benchmarking, data centers, or chip sales.
- Use the interactive chart to filter by developer, country, or date, then read the notes under it for the method.
- Download the CSV, or install the epochai Python client if you need the relationships between models and benchmark runs intact.
- For an evaluation or a custom report, send a note through the consultations page; those engagements are paid and reviewed for fit.
Product Information
A quick look at Epoch AI's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- AI policy analysts
- Machine learning researchers
- Journalists and editors
Tasks
- Checking a training compute estimate before quoting it
- Comparing model progress across benchmarks
- Sizing the power footprint of AI infrastructure
Scenarios
- A team drafting an internal AI strategy memo on a deadline
- A newsroom verifying a vendor benchmark claim on launch day
- A lab that needs an outside evaluation of a new model
Key features
FrontierMath Benchmark
FrontierMath tests models on original problems written by mathematicians, split into four tiers that run from advanced undergraduate work to short research projects. Problems stay unpublished and answers are machine-checkable, so a high score is hard to fake through memorization. The program also includes Open Problems, a set of unsolved research questions, and a collection of Erdős problems formalized in Lean.
Epoch Capabilities Index
The Epoch Capabilities Index, usually shortened to ECI, rolls many benchmark results into a single number so you can compare models that were never tested on the same set. It's built from real observed runs and fitted across benchmarks, which makes it useful for tracking the frontier over time. Treat the raw value as abstract and read the benchmark curves underneath when you need something concrete.
AI Models Database
This is the core dataset: more than 3,200 machine learning models from 1950 to today, with training compute, developer, publication date, and access type. Each training compute figure comes with an estimate and a confidence note rather than a bare number. For anyone studying scaling, it saves months of manual collection.
Benchmarking Hub With API Access
Epoch AI runs its own evaluations and also collects results published by others, keeping the two in separate tables so you can see which is which. The data comes as CSV, as browsable tables, or through the epochai Python client, which reads an Airtable API and keeps links between models and runs intact. That last part matters when you need to join a score to a specific model version.
AI Data Center and Hardware Data
The AI data center database maps large facilities using satellite imagery and permit filings, and estimates their compute and power draw. Related datasets cover machine learning hardware, GPU clusters, chip sales, and company investment. Together they answer questions that model leaderboards ignore, like where the electricity is going.
Commissioned Research and Evaluations
Companies, nonprofits, and governments can pay Epoch AI to run model evaluations, build a custom dataset, or write a report. Past work includes contributions to Stanford's AI Index and the UK's Department for Science, Innovation and Technology. It's a consulting arrangement with a public-interest filter, not a self-serve service.
Pros and cons
Pros
- The datasets are free to download and openly licensed, which is rare for AI infrastructure numbers.
- FrontierMath resists contamination because the problem set stays private and answers can be verified by machine.
- The AI models database goes back to 1950, so long-run trend analysis doesn't require stitching sources together.
- The epochai client keeps relationships between models and benchmark runs, which plain CSV exports tend to flatten.
Cons
- Updates arrive in batches, so a model released this week may be missing from the AI models database for a while. Plan around the lag if you publish on launch day.
- Scores can differ from the numbers a lab reports, and the notes explain why, but you'll need to read them before citing either figure.
- The API runs on Airtable, which brings rate limits and key management that a purpose-built API would handle for you.
- Commissioned work has no public rate card and depends on fit, so budgeting starts with a conversation rather than a price page.
Frequently asked questions
Epoch AI studies how AI is developing and publishes the underlying data. That means databases of models, chips and data centers, independent benchmark runs, and reports on scaling, economics, and impact.
Related content
Explore related tools, skills, and articles for Epoch AI.
Epoch AI Alternatives
Google EmbeddingGemma 2
Google DeepMind · Writing · Image · CodingGoogle EmbeddingGemma 2 is a 740M-parameter open model that turns text, code, images, audio, and video frames into vectors in one 768-dimension space, running on phones and laptops.
MePersona
MePersona, Inc. · Writing · Productivity · MarketingMePersona is an AI unified inbox for creators. It learns your voice, drafts replies to fan messages from 12 platforms, and sends only what you approve.
Fluence
Yay Software (Yannick Huchard) · Writing · Voice & Language · ProductivityFluence is an Android app that reads PDFs, EPUBs, Word files and web articles aloud, follows each sentence as it speaks, and answers questions about the text on the phone itself.