GameBench 2 — Current standings
Overall Models
AI-written players, ranked by match results. This view equally averages each model's eligible reasoning settings across the public games. Each attempt whose player program could not be generated counts as 0.
- Updated
- 2026-10-26
- Model families
- 97
- Reasoning settings
- 261
- Player programs
- 2,053
- Public games
- 8
Filtered view
Filters change this view; published ranks remain in the Official column.
Rankings
Models
Overall leaderboard
How to read
We ask each AI model to write a player program for every game. DuelLab runs those programs against each other. Higher scores mean stronger results against a fixed reference field of player programs.
Score is the 0–100 result across public games: expected points against the fixed reference field. Working-program strength averages only games where the model produced a working player program, leaving failed generation attempts out. Codegen shows the share that worked first try / worked after repair or resend / failed. Players counts the player programs behind a row, while Best is its highest single-game score. W / D / L shows rated wins, draws, and losses. More matches with existing programs do not add independent generation evidence. Where shown, the reasoning spread tells how much results vary across settings. Opponents played, hidden until you turn it on under Columns, shows how far the row's player programs are through the full round-robin of their games.
Partial 3/3 is an official Overall rank whose evidence is incomplete in another way, so compare it with care. Models with results for fewer than three reasoning settings get no Overall rank. Provisional rows do not receive an official rank. Capped draws reached the shared move cap and still count as draws. Filters can change the visible order, but the Official or Base column keeps the published rank. Cost / player is an estimate; select a value for its source and pricing details.
Columns
Choose the optional columns shown in this table.
Models
Not ranked Overall (31)
An Overall score averages a model's three reasoning settings. These models have results for fewer settings, or still miss evidence for one, so they are listed here without an Overall rank. Their settings are ranked in the Model + reasoning setting view. Some have more than one reason.
- 31Fewer than three reasoning settings An Overall score averages three reasoning settings; an average of fewer is not comparable. Each setting is ranked under Model + reasoning setting.
- 2Not every game attempted Some games have no player-program attempt: their generation is still running or stopped before them. The score so far covers only the games attempted.
- 6Some games cannot be scored A game's only generation attempts failed because of our test harness or the provider, not the model, and cannot currently be repeated. The score averages the other games.
- 4Too many programs failed Too few of its player programs compiled and ran to qualify for a rank.
Show all 31 entriesHide entries
| Model | Settings | Average of its settings | Why not ranked yet |
|---|---|---|---|
 Pareto 26.10 | 1/3 | 90.6 | Fewer than three reasoning settings |
| Pareto 26.10 Preview | 1/3 | 89.7 | Fewer than three reasoning settings |
 MiMo-V2.6-Pro | 2/3 | 62.1 | Fewer than three reasoning settings |
 Step 5 Preview | 2/3 | 58.6 | Fewer than three reasoning settings; Not every game attempted |
| MiMo-V2.6-Flash | 2/3 | 50.4 | Fewer than three reasoning settings |
 Qwen3.8 Flash | 2/3 | 48.4 | Fewer than three reasoning settings |
 Mistral Large 4 | 2/3 | 46.7 | Fewer than three reasoning settings |
 Glimpse | 2/3 | 42.7 | Fewer than three reasoning settings |
 GPT-4.1 | 1/3 | 42.4 | Fewer than three reasoning settings |
 Kimi K2.7 Code | 1/3 | 40.6 | Fewer than three reasoning settings; Some games cannot be scored |
| GPT-5.4 Mini | 2/3 | 38.9 | Fewer than three reasoning settings |
| MiMo-V2.5-Pro | 2/3 | 38.3 | Fewer than three reasoning settings; Some games cannot be scored |
| Qwen3.7 Max | 2/3 | 37.6 | Fewer than three reasoning settings |
 Grok Build 0.1 | 1/3 | 34.9 | Fewer than three reasoning settings |
 Ling 3.1 Flash | 1/3 | 33.3 | Fewer than three reasoning settings; Not every game attempted |
 Laguna S 2.1 | 2/3 | 31.6 | Fewer than three reasoning settings |
 Hy3 Preview | 2/3 | 31.4 | Fewer than three reasoning settings |
 Nemotron 3.5 Super Vl | 2/3 | 30.9 | Fewer than three reasoning settings; Too many programs failed |
| MiMo-V2.5 | 2/3 | 30.7 | Fewer than three reasoning settings; Some games cannot be scored |
| GPT-5.4 Nano | 2/3 | 29.5 | Fewer than three reasoning settings |
 Gemma 4 31B | 2/3 | 29.2 | Fewer than three reasoning settings |
 LongCat 2.0 | 2/3 | 28.2 | Fewer than three reasoning settings |
 Minimax M3 | 2/3 | 26.7 | Fewer than three reasoning settings |
 Nex N2 Pro | 2/3 | 26.5 | Fewer than three reasoning settings; Some games cannot be scored |
| Mistral Medium 3.5 | 2/3 | 25.9 | Fewer than three reasoning settings |
| Qwen3.7 Plus | 2/3 | 25.6 | Fewer than three reasoning settings |
| Ling 3.0 Flash | 2/3 | 24.7 | Fewer than three reasoning settings; Some games cannot be scored |
| Flint V1.0 | 1/3 | 23.6 | Fewer than three reasoning settings |
| Nex N2.5 Pro | 2/3 | 19.6 | Fewer than three reasoning settings; Some games cannot be scored; Too many programs failed |
 North Mini Code | 2/3 | 15.0 | Fewer than three reasoning settings; Too many programs failed |
| GPT-4.1 Nano | 1/3 | 11.8 | Fewer than three reasoning settings; Too many programs failed |
Scoring method
How this is scored
Each model is asked to create a program that can play every benchmark game.
The generated programs compete head-to-head, with both players receiving comparable opportunities.
Wins, draws, and losses produce a B1 (Bradley–Terry) rating for each program, which becomes a 0–100 score for each game: 100 × its expected points against the fixed reference field.
The main score is the mean across games. Every generation attempt counts: a game combines its working programs' scores with a 0 for each attempt whose player program could not be generated, and a reasoning setting where every attempt failed scores 0. A program rejected only for being too slow, or lost to a fault of the test machine, does not play and does not count. The share of working programs is shown on its own (Codegen).
The reasoning settings are combined into one row for each model.
Detailed tables provide match records, match counts, and other context for careful comparison.