for some years now, I like to test llm against minesweeper, and I am always disapointed because they used to always fail, until today.
Today I tested Claude code with the following board, a simple 8×8 with 9 mines and it nailed, in max effort opus 4.8.
the next test was a 9×9 grid, and it almost nailed on max and opus 4.8
this was after the first iteration, and in the second it misplaced two flags.
in ths one it ran out of context tho.
also tried a 16×16 board on 4.8 with high effort and also did good, this is the first iteration that claude did, x is bomb, check is safe to clic
this are 3 iterations, at the end the board looked like this. this iteration is interesting because it wrote a python solver for minesweeper. even though it says it ran the logic by “hand” and use the solver just to validate.
and by now, board is almost solved.
last iteration
My takes
It solve the game in 5 iterations in opus 4.8 in high, Is interesting to me that opus 4.8 max failed in a smaller board. overall impressive that it could solve it. First time I see an llm doing it.
small table summary
| model | effort | board | result |
| claude opus 4.8 | max | 8×8 | success |
| claude opus 4.8 | max | 9×9 | fail |
| claude opus 4.8 | high | 16×16 | success |