Benchmark Almanac
New Terminal-Bench 4.1, Speech Agent Arena and 1,540 more

Benchmark Almanac

We track 3,322 benchmarks from 1,746 public leaderboards, 522 launch posts, 346 system cards, and 1,233 open-source GitHub repositories.

Click on any of the benchmarks to open them!

Catching the frontier

We rank models based on Epoch AI's Capabilities Index, an aggregate score from many benchmarks.

All models
Anthropic Nov 2026

Claude Fable 5.2

Closed weights Flagship

173.4 #1 of 256
Epoch Capabilities Index
Self-Reported Index 175.3 #1 of 404
  • AutomationBench 51.5% #2 /42
  • GPQA Diamond 93.9% #18 /229
  • HealthBench Professional 67.7% #3 /26
Google DeepMind Oct 2026

Gemini 4 Carbon

Closed weights

169.8 #2 of 256
Epoch Capabilities Index
Self-Reported Index 173.2 #2 of 404
  • AutomationBench 53.9% #1 /42
  • GPQA Diamond 96.5% #1 /229
  • FrontierCode 2.0 42.6% #4 /32
OpenAI Sep 2026

GPT-6 Astra

Closed weights Flagship

169.0 #3 of 256
Epoch Capabilities Index
Self-Reported Index 172.1 #3 of 404
  • Humanity's Last Exam 54.8% #1 /47
  • DeepSWE 74.1% #1 /28
  • AutomationBench 41.4% #7 /42
Anthropic Sep 2026

Claude Opus 5.5

Closed weights Flagship

168.5 #4 of 256
Epoch Capabilities Index
Self-Reported Index 170.0 #5 of 404
  • AutomationBench 42.5% #6 /42
  • Agents' Last Exam 38.2% #1 /31
  • Terminal-Bench 4.0 64.8% #1 /40
OpenAI Sep 2026

GPT-6.1 Sol

Closed weights Mid-size

168.5 #5 of 256
Epoch Capabilities Index
Self-Reported Index 164.0 #15 of 404
  • MultiChallenge 82% #1 /42
  • Terminal-Bench 4.0 58.2% #5 /40
  • GPQA Diamond 95.4% #6 /229
OpenAI Oct 2026

GPT-6.1 Terra

Closed weights

167.7 #6 of 256
Epoch Capabilities Index
Self-Reported Index 159.8 #32 of 404
  • MultiChallenge 74.2% #7 /42
  • Terminal-Bench 4.0 51.8% #8 /40
  • GPQA Diamond 94.3% #11 /229

Who is dominating the leaderboards?

We track the 7 most-cited benchmarks in model releases in the past year. Terminal-Bench was the most cited, featured in 144 releases.

Every category
Google DeepMind Anthropic OpenAI xAI DeepSeek Moonshot AI MiniMax Qwen Xiaomi Corp NVIDIA Inception Labs Meta
Terminal-Bench 4.0 19 leaders
Claude Opus 5.5 64.8%
GPQA Diamond Passed 85% in Jun 2025, by Gemini 2.5 Pro
Gemini 4 Carbon 96.5%
Humanity's Last Exam 12 leaders
GPT-6 Astra 54.8%
SWE-bench Verified Not tested since Jun 2026
Claude Opus 4.7 83.5%
AIME 2026 Passed 85% in Feb 2026, by GPT-5.2
GPT-5.5 100%
tau3-bench Passed 85% in Jun 2026, by Nemotron 3 Ultra
Mercury 2.5 96%
SWE-Bench Pro Replaced by V2, Sep 2026
Muse Spark 1.1 61.5%

Sources: Official leaderboards, Epoch AI runs (CC BY 4.0), boards copied by Epoch AI (CC BY 4.0) and labs’ model cards

Google DeepMind Anthropic OpenAI xAI Z.ai DeepSeek Moonshot AI Qwen Meta
Terminal-Bench 4.0 19 leaders
Claude Opus 5.5 64.8%
GPQA Diamond Passed 85% in Jun 2025, by Gemini 2.5 Pro
Gemini 4 Carbon 96.5%
Humanity's Last Exam 12 leaders
GPT-6 Astra 54.8%
SWE-bench Verified Not tested since Jun 2026
Claude Opus 4.7 83.5%
AIME 2026 Passed 85% in Feb 2026, by GPT-5.2
GPT-5.5 100%
tau3-bench Passed 85% in Mar 2026, by Qwen3.5 397B-A17B
Qwen3.5 397B-A17B 87.9%
SWE-Bench Pro Replaced by V2, Sep 2026
Muse Spark 1.1 61.5%

Sources: Official leaderboards, Epoch AI runs (CC BY 4.0) and boards copied by Epoch AI (CC BY 4.0)

Anthropic OpenAI Google DeepMind Cognition xAI Moonshot AI ByteDance Seed DeepSeek Qwen MiniMax Inception Labs
Terminal-Bench 4.0 20 leaders
Claude Sonnet 5.5 70.6%
GPQA Diamond Passed 85% in Jun 2025, by Gemini 2.5 Pro
Gemini 4 Carbon 96.5%
Humanity's Last Exam 15 leaders
Claude Fable 5.2 74.2%
SWE-bench Verified Passed 85% in Apr 2026, by Claude Mythos Preview
Claude Opus 5 96%
AIME 2026 Passed 85% in Feb 2026, by GPT-5.2
GPT-5.5 100%
tau3-bench Passed 85% in Mar 2026, by Qwen3.5 397B-A17B
Mercury 2.5 96%
SWE-Bench Pro Passed 85% in Sep 2026, by Claude Opus 5.5
Claude Fable 5.2 93%

Sources: Official leaderboards, Epoch AI runs (CC BY 4.0) and labs’ model cards

Leaderboards to watch

Here are the six most popular benchmarks that haven't yet saturated (score >85%).

All benchmarks in use

What do labs feature the most?

Agent tests rose from 2% of what labs headlined in 2024 to 19% in 2026, as knowledge tests fell from 52% to 19%.

All model cards
EverydaySchoolUniversityCompetitionExpertResearch frontier

Select an area to break it down.

0%25%50%75%100%Knowledge and languageMath and scienceCodingAgents and toolsProfessional workVision, audio and roboticsSafety and security20232024202520260%25%50%75%100%2023202420252026

Source: Labs’ launch posts, read by Benchmark Almanac

Saturation is one reason labs drop a benchmark

22 benchmarks appeared in five or more launch posts and then went a year without one. Labs dropped 10 of them after the best score passed 85%. Below are the 36 benchmarks launch posts list most often, in the order labs took them up. Each tick is a post.

MMLU HumanEval GSM8K AI2D MATH BIG-Bench Hard MBPP IFEval MMMU MathVista GPQA Diamond LiveCodeBench MMLU-Pro MMMLU SWE-bench Verified AIME 2024 MATH-500 SimpleQA MATH-Vision (MathVision) MMMU-Pro Video-MME Humanity's Last Exam AIME 2025 CharXiv BrowseComp LiveCodeBench v6 HMMT 2025 IFBench tau2-bench SWE-bench Multilingual SWE-Bench Pro Terminal-Bench 2.0 MCP Atlas Terminal-Bench 2.1 AutomationBench DeepSWE
In use: two or more cards in the last six months Fading Dropped: no card for a year Best score passed 85% Best score reached 95%

Source: Labs’ launch posts, read by Benchmark Almanac

Found this week

New benchmarks our discovery job found in paper feeds, newsletters and registries. They're queued for the catalog.

Everything new

What gets in

We add a benchmark if anyone can run it again and it passes at least one of these four tests.

How we decide
  1. A major lab reports it A lab that ships frontier or widely used models reports it in a launch post, model card, system card or technical report.
  2. An independent leaderboard runs it Its makers or a third party keep a leaderboard with scores for five or more current models.
  3. An established evaluation group made it For example Epoch AI, METR, Scale, Vals, ARC Prize, Apollo Research, UK AISI, US CAISI, SecureBio, Arena, HELM or Kaggle.
  4. People use it About 100 GitHub stars or 30 Hugging Face upvotes, coverage in a widely read newsletter, or other papers and model cards that run it.

Test sets made only for one paper’s own method, training datasets and renamed copies stay out. The New page lists every find we left out, with the reason.

How to contribute

The almanac is open source. Every change goes through GitHub, and a maintainer checks its source before it lands.

Open tasks
  • Suggest a benchmark Send it through the submission form, with a link to its paper, code or leaderboard.
  • Add a score or fix a record The Contribute page lists open tasks. Each one opens a form that names what’s missing.
  • Point us at your results If you made a benchmark, tell us where its results live and we’ll import them.
  • Send code New leaderboard importers and fixes go straight to a pull request.