MINES BENCHMARK · LOCAL, THINKING LOW

gpt-oss 20B

Openai · open weights · released 2025-08-05 · route gpt-oss:20b · Model profile · Chess results · Sudoku results · FreeCell results · Nations Royale results · Hustle Sim results · Elemental Armory results · Micro Mouse results

Results

Score
095%: 0 to 27 · 4 of 24 base boards so far
Boards won
0 of 4
Safe cells cleared beyond the opening
20.0%
Won / hit a mine / unfinished / unreadable
0 / 4 / 0 / 0
Ended for changing nothing five turns running
0
Reached the turn cap
0
Flags left on safe cells
3 in 1 boards
Illegal actions
2 of 73 (2.7%)
Wasted actions (already open or flagged)
9 of 73 (12.3%)
Unreadable replies
0
Turns per board
1.8
Requests per board
1.8
API-equivalent cost per win
n/a
Tokens per board (in / out)
1,748 / 25,733
Reasoning tokens per board
n/a
API-equivalent cost per board
n/a
Seconds per board
2,132.4
Replies cut off at the length limit
0.0%
Aborted on our side (not counted)
5

By tier

TierBoard, minesWonClearedTurnsOutput tokensSeconds
Beginner9×9, 100 of 173%4.061,3971,584.0
Intermediate16×16, 400 of 17%1.023,5403,286.6
Expert30×16, 990 of 10%1.09,7072,955.5
Master30×24, 2000 of 10%1.08,288703.5

Cells by the technique that proves them

Each safe cell covered at the start is tagged with the simplest deduction that first proves it safe, given everything a perfect player could open before it.

TechniqueCellsOpened
Cascade from an empty cell4588.1%
One number5693.0%
Two numbers together685.9%
The mine count40.0%

Every board