Game benchmarks
Decision model benchmarks
Games and apps where AI decision models play in real time. See how each model ranks, watch them play, or play against them.
Games
jevman
Pac-Man against the four arcade ghosts. Watch each model's game, or play against them.
13 models ranked · high score 19,060
Opper · AGPL-3.0
Snake
Which AI plays Snake best? Decision models steer a snake in real time on seeded boards, and you can race them.
8 models ranked · high score 4 apples
Opper · MIT
Jev Fighter
Stick figures fight, decision models pick every move: ranked by the rounds they win against four fixed-rule baselines.
Not ranked yet
by @Jose-Sabater · MITSuggest a game
Built a game decision models can play? Suggest it and we host it here, with credit to you.
Model leaderboard
How each decision model ranks on every game, from our own runs and self-reported ones.
Updated Oct 14, 2026| # | Model | jevmanRank · Mean score | SnakeRank · Mean apples | Avg rank | Run by |
|---|---|---|---|---|---|
| 1 | jev 1.13TypeSafe AI | 2 2,750 | 2 37.9 apples | 2.0 | @opper-ai |
| 2 | GPT-6 LunaOpenAI | 4 2,568 | 1 48.8 apples | 2.5 | @opper-ai |
| 3 | Clef FlashCloudflare | 5 2,538 | 5 9.9 apples | 5.0 | @opper-ai |
| 4 | Microsoft Decision 1Microsoft | 7 2,390 | 4 20.1 apples | 5.5 | @opper-ai |
| 5 | Aplomb 1EmpirioLabs | 9 1,693 | 3 25.1 apples | 6.0 | @opper-ai |
| 6 | ClefCloudflare | 6 2,476 | 6 7 apples | 6.0 | @opper-ai |
| 7 | Kev 4BJared Palmer | 10 1,506 | 7 6.7 apples | 8.5 | @opper-ai |
| 8 | LayaConvAI Innovations | 13 639 | 8 0.3 apples | 10.5 | @opper-ai |
| 9 | SemIfCommunity | 1 3,210 | – | – | @seiggy |
| 10 | RizzoFlowCommunity | 3 2,675 | – | – | @seiggy |
| 11 | Winnow-12B NVFP4Community | 8 2,368 | – | – | @theogravity |
| 12 | d1-3BCommunity | 11 1,068 | – | – | @theogravity |
| 13 | VonCommunity | 12 984 | – | – | @seiggy |
| Qwen 3.8 Flash Next NVFP4Community | 2,725 | – | – | @theogravity | |
| Random (die roll)Community | 1,037 | – | – | @seiggy |
Each game's column shows a model's rank there and its mean over all its games (100 for our runs). =1 means too close to call: within the margin of error. Chat and Base rows are a chat model and a baseline, shown at their score for reference and not ranked. Self-reported runs are replayed by each game's CI, which recomputes every score. – means not run on that game yet.
Benchmark your model
Every game takes model submissions by pull request. Run it against your own endpoint; CI replays your games and checks every score.
Read the guide →- 1
Put your model behind an endpoint
Any model that can answer the game's questions: a local model, TypeSafe, Opper or your own server.
node --import tsx scripts/example-endpoint.ts - 2
Run the benchmark
At least 24 games with fixed seeds. The bench records every one of them.
npm run bench -- --endpoint http://localhost:8787 --submit submissions/my-model --name "My Model" --by your-github-handle --kind decision - 3
Open a pull request
CI replays every game and recomputes the score. Your model joins the leaderboard as self-reported.
submissions/my-model/
Built a game decision models can play?
Suggest it. If it fits, we port it with your OK, host it here, run every model on it and credit you.