DuelLabGameBench 2
Menu

Model results · OpenAI

GPT-OSS 120B

Rank 59 of 66 in Overall Models. All three benchmark settings tested.

Score · relative 0–100
30.2
Rank
59
Status
Complete 3/3
Rank change
0
Players generated
2026-09-03
Match evidence updated
2026-10-26
Results published
Compare in charts View player programs Back to leaderboard

Scores by reasoning setting

  • High (Intensive group, tested and included)35.1
  • Medium (Balanced group, tested and included)31.2
  • Low (Baseline group, tested and included)24.3

Every tested setting is included in the overall score.

Available settings and grouping

Available settings

LowMediumHigh

Checked on 2026-07-30 · Provider documentation

How settings are grouped

  • Baseline: Low — tested and included
  • Balanced: Medium — tested and included
  • Intensive: High — tested and included

Scores by game

  • Go1 of 3 failed25.7
  • Chess43.2
  • Shogi39.6
  • Xiangqi1 of 3 failed27.2
  • Hex29.9
  • International Draughts1 of 3 failed16.5
  • Game of the Amazons28.1
  • Tumbleweed31.7

Each game counts every generation attempt, a failed one as 0 (“1 of 2 failed” marks games with failed attempts). Scores are relative within each game, across the included settings; they do not compare game difficulty.

What this result shows

Strongest included setting: High (35.1).

Across included settings, the highest game score is Chess (43.2); the lowest is International Draughts (16.5). Each game counts every generation attempt, a failed one as 0. These are relative scores within each game, not direct comparisons of difficulty.

Match evidence measures these retained programs. It does not establish how a fresh generation will perform. Independent-generation variation is not estimated in this release.

Test protocol and evidence limits
Model identifier
openai/gpt-oss-120b
Provider route (catalog)
openrouter
Prompt identity / profile
Not retained in this result set / Not retained in this result set
Prompt version
Not retained in this result set
Generation policy
Generation-quality admission policy (see methodology)
Requested output-token limits
Not retained (0 of 33 indexed attempts)

Generation evidence

  • Low: 6 contributing player programs across 6 games
  • Medium: 8 contributing player programs across 8 games
  • High: 7 contributing player programs across 7 games

Counts describe contributing programs, not independent replications: repairs and resends are not new independent samples. Independent-generation lineage is not retained.

GameReasoning groupProgramsWorking-program strengthRaw ratingCombined record (W / D / L)
GoBalanced1 working of 1 attempt36.01333.7129 / 94 / 293
GoBaseline1 working of 1 attempt41.01383.3159 / 85 / 272
ChessIntensive1 working of 1 attempt43.91424.312 / 381 / 87
ChessBalanced1 working of 1 attempt42.51413.75 / 382 / 93
ChessBaseline1 working of 1 attempt43.31420.31 / 398 / 81
ShogiIntensive1 working of 1 attempt46.41396.2151 / 118 / 225
ShogiBalanced1 working of 1 attempt35.91304.599 / 122 / 273
ShogiBaseline1 working of 1 attempt36.31308.3103 / 118 / 273
XiangqiIntensive1 working of 1 attempt42.51417.140 / 327 / 131
XiangqiBalanced1 working of 1 attempt38.91389.022 / 328 / 148
HexIntensive1 working of 1 attempt58.31542.7300 / 0 / 230
HexBalanced1 working of 1 attempt15.7926.078 / 0 / 452
HexBaseline1 working of 1 attempt15.7926.078 / 0 / 452
International DraughtsIntensive1 working of 1 attempt22.91007.597 / 2 / 403
International DraughtsBalanced1 working of 1 attempt26.61065.9112 / 6 / 384
Game of the AmazonsIntensive1 working of 1 attempt34.11247.2167 / 0 / 363
Game of the AmazonsBalanced1 working of 1 attempt21.21063.0101 / 0 / 429
Game of the AmazonsBaseline1 working of 1 attempt28.91177.4140 / 0 / 390
TumbleweedIntensive1 working of 1 attempt32.61199.0103 / 118 / 317
TumbleweedBalanced1 working of 1 attempt33.01203.240 / 248 / 250
TumbleweedBaseline1 working of 1 attempt29.41163.434 / 223 / 281

Each row covers all the model's programs in one game and setting. Strength and raw rating are attempt-weighted means; under them, each program's score and the lowest and highest rating describe the programs (not a confidence interval). The combined record adds up their matches: evidence volume, not strength.

Failures and budgets

3 model-code generation failures; 0 provider/infrastructure failures excluded from the generation denominator. Runtime player faults are separate: 0 recorded one-sided faults across contributing program rows. A crash or illegal move pinned on one program also counts as that program's loss in its match record; other faults, such as timeouts, are left out of rating evidence.

Repair/resend outcomes and available costs are included. Exact retry ceilings and execution budgets are not retained in this result set; see the methodology for the public policy. Do not infer a measured setting from a current provider catalog.

Player program results

High: 63% / 25% / 12% · n 8Medium: 75% / 25% / 0% · n 8Low: 63% / 12% / 25% · n 8 All three benchmark settings testedMany programs failed

Generation cost

$0.0033average estimated cost per game and reasoning setting

Median
$0.0014
Range
$0.000467–$0.02
Provider-reported total
$0.08
Cost data
24 combinations · 24 runs
Price source
Provider-reported
Price list
Not recorded for this older data
Average output
9.3k tokens
Output range
1.5k–64k tokens
Runs with output-token data
24
How cost is measured

Comparison cost gives equal weight to each model, game, and reasoning-setting combination. Recorded generation attempts include repairs and resends when cost evidence is available; runtime compute and service operation are excluded. Cost and output-token counts use separate telemetry denominators. A list-price estimate is not actual expenditure; mixed evidence is not a uniform standard-price comparison.

Test protocol

  • Combines this model's tested reasoning settings
  • Player-program generation; retained protocol details are under “Test protocol and evidence limits” above