Benchmark Almanac

I wanted an easy way to see which benchmarks and arenas already exist, how models perform on them, and how evals have changed over time. So I built Benchmark Almanac, which tracks 3,322 benchmarks from 1,746 leaderboards, 522 launch posts, 346 system cards and 1,233 open-source GitHub repos. I launched it on November 13, and it is fully open source.

What do labs report?

Labs named 875 benchmarks in their 2026 launch posts and system cards, but they’re spread out: 570 were named by just one lab. And it turns over fast. Only GPQA Diamond made the top ten in every quarter since 2024.

Hover a shape for its benchmark.

Have evals gotten harder?

Much harder. Expert and research-frontier tests went from 5% of what labs report in launch posts in 2024 to 38% in 2026. And 2026 was the year of agents: agent tasks (25%) passed multiple choice (20%), which was 46% back in 2023.

Hover a half year for its numbers.

Except for robotics

Expert-level benchmarks dominate professional work (84%), but only 4% of vision, audio and robotics evals can be considered expert-level. 96% of the robotics tests labs headlined this year are everyday tasks most adults can do.

Each bar is one area’s tests, easiest to hardest.

Which lab leads where?

Evals also offer a window into a model’s jagged capabilities. Anthropic leads 10 of the 13 contested bio and chem benchmarks, while OpenAI models held number one in cyber 57% of the time and in math 55%. Robotics is still heavily contested.

Hover a bar for the lab.

Build your own index

If you care about a blend of benchmarks more than any one number, you can make your own: add or drop evals and give each one your own weight. The six starter mixes crown five different models.

Drag a weight.

Compare models head to head

You can also put up to four models head to head on every benchmark they’ve all been evaluated on. On launch day, Claude Fable 5.2 beat Gemini 4 Carbon, GPT-6 Astra and Muse Wonder on 21 of the 45 benchmarks all four share.

Pick two to four models. Each square is a benchmark.

It stays current

The almanac is meant to stay current. Every morning, one job reads 26 sources for new benchmarks and models, and opens a pull request for maintainers to review.

What people built with it

The day after launch, Ethan Mollick used the almanac’s data to build SatBench, which treats the 173 benchmarks labs cite most as one big exam. Perceptron sent the first outside pull request, adding its EgoHome benchmark.

The almanac’s own charts on this page use its data from launch day, November 13th, 2026, with Epoch AI’s scores under CC BY 4.0. The live site has more by now.