Model rankings from blind human votes
We rank models by blind human preference on real tasks: models never know they are competing, votes are quality-weighted into Bradley-Terry ratings with 95% confidence intervals, and the raw votes are published nightly — every number here can be recomputed from the dump. Published benchmarks, which tend to find their way into training data, appear alongside as provenance-tracked context; our ratings remain the verdict.
Current rankings
Bradley-Terry ratings with 95% confidence intervals from eligible blind votes. Overlapping intervals share a rank band; provisional models are rated but unranked.
No official snapshot yet — ratings are recomputed every six hours once eligible votes exist.
The unified boards also show published benchmark scores as context, each with full provenance — arena ratings from blind human votes remain the primary signal.
Beyond the ranking
What the ratings look like against cost, which models moved, and who joined the roster.
Rating vs cost per response
All charts →Higher and further left is better.
Appears once a rating snapshot and generation telemetry exist.
Biggest movers
History →Rating change between the two latest snapshots.
Movers compare the two most recent snapshots — they appear once a category has rating history.
Newest models
All models →Latest additions to the active roster.
- Llama 4 MaverickMeta · 2025-04-05
- GPT-5.6 LunaOpenAI · 2026-07-09
- DeepSeek V4 Pro (0813)DeepSeek · 2026-08-13
- Qwen3.8 Max (0902)Alibaba Qwen · 2026-09-02
- Grok 4.6SpaceXAI · 2026-08-12
- Gemini 3.8 FlashGoogle · 2026-09-02
Blind battles, weighted votes, reproducible ratings
Every ranking on this site traces back to a human choosing between two anonymous responses — and to data you can download and check.
Blind battle
Two anonymous models answer the same real task. Blinding is structural: contestants are addressed by slot only, so a model never knows it is competing — there is nothing to game.
Human vote
A person compares both responses side by side and votes before identities are revealed. Arena creators judge their own battles; the judging queue collects votes from many people.
Weighted ratings
Votes are quality-weighted and fed into Bradley-Terry ratings with bootstrap 95% confidence intervals. Overlapping intervals share a rank band; new models stay provisional until they have enough votes.
Published nightly
Raw votes and battles are dumped nightly in a documented format. Run the open ratings engine on the dump and you get the same numbers this site shows.
The rankings are made of your votes
Every vote is cast blind and lands in the public dump.
Enter the Arena
Pick a category, bring your own prompt or use a curated one, and vote blind on two live responses.
Start a battle →Judge the queue
Vote on finished battles from other people's prompts. More judges per battle means tighter confidence intervals.
Open the queue →Suggest a model
Missing a model you care about? Nominate it for the roster and we'll review it.
Suggest a model →What changed
Every rating snapshot, roster change and article, in the open.
Changelog
Full changelog →- new modelModel added: Kimi K32026-09-08 21:23 UTC
- new modelModel added: Qwen3.8 Max (0902)2026-09-08 21:23 UTC
- new modelModel added: Mock Beta2026-09-08 21:23 UTC
- new modelModel added: Llama 4 Maverick2026-09-08 21:23 UTC
- new modelModel added: Grok 4.62026-09-08 21:23 UTC
- new modelModel added: DeepSeek V4 Pro (0813)2026-09-08 21:23 UTC
- updateBrowser scene promoted: c3b-walk vetted scene2026-09-08 02:17 UTC
- new modelModel added: GPT-5.6 Luna2026-09-07 20:15 UTC
- new modelModel added: Gemini 3.8 Flash2026-09-07 20:15 UTC
- new modelModel added: Claude Sonnet 52026-09-07 20:15 UTC
- new modelModel added: GPT-5.6 Terra2026-09-07 20:15 UTC
- new modelModel added: Claude Opus 52026-09-07 20:15 UTC
Latest research
All articles →- How HumanEval.org ratings workHumanEval.org team · 2026-08-23 21:56 UTC
Stay in the loop
Launch notes, new models and methodology changes — straight to your inbox, nothing else.