We track 2,721 benchmarks: what each one tests, which model is ahead, and whether labs still report it. We take scores from public leaderboards, Epoch AI, and 479 launch posts and 320 system cards from 51 labs.
We drew each benchmark's best score since 2023 as one line, colored by area. Click a line to open it.
Ranked by Epoch AI's Capabilities Index, one score built from many benchmarks.
Closed weights Flagship
Closed weights Flagship
Closed weights
Closed weights Mid-size
Closed weights Mid-size
Closed weights Flagship
The capability index and six of the most-used math benchmarks, since 2023. Taller means a higher score.
No model has reached 85% on these yet, and labs have reported them in launch posts over the last six months.
113 original long-horizon engineering tasks across 91 repos and five languages, written from scratch. Hand-written verifiers accept alternative solutions.
Gemini 3.8 Flash 73.8% Official leaderboard
2,500 expert-written closed-ended questions across 100+ subjects. Exact-match or multiple-choice answers graded by an LLM judge.
Gemini 3.1 Pro 46.4% Gemini 3.8 Flash 44.5% Official leaderboard
Agent completes end-to-end business workflows inside a simulated company exposing dozens of REST APIs. Graded on resulting state.
Gemini 4 Carbon 53.9% Gemini 4 Argon 51.3% Official leaderboard
1,865 long-horizon issue-resolution tasks from 41 repos in four languages. Agent patch must pass hidden tests. Pass@1.
Muse Spark 1.1 61.5% Muse Spark 55% Gemini 3.1 Pro 46.1% Official leaderboard
1,500+ expert-sourced tasks from real projects across 55 occupations in three difficulty tiers. Objective verifiers grade deliverables. 5-hour cap.
Gemini 4 Carbon 45.9% Official leaderboard
58 new out-of-domain verifiable output constraints. Scored by programmatic constraint checks (strict or loose).
Nemotron-Cascade 2 30B-A3B 82.9% 
Qwen 3.8 Max 82.8% 
Inkling-Small 82.2% Nemotron 3 Ultra 81.7% Qwen3.8-Omni-Flash 81.5% As labs report it
Each area's share of headline benchmarks in launch posts, by half year.
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.
New benchmarks our discovery job found in paper feeds, newsletters and registries. They're queued for the catalog.
Cogito v2 Preview 671B MoE 
DeepSeek-Coder-V2 236B