Thesis
A visual game benchmark measures an agent system, not just the model, because success depends on perception, action selection, and persistence across a long task.
Observed facts
Epoch AI's SpireClimb evaluates models in Slay the Spire 1 and Slay the Spire 2 using a single-agent visual computer-use harness.
1The report says frontier systems exceed 90% win rates on the original game's base difficulty, while GPT-6 Astra and GPT-6.1 Sol reach 17% and 20% respectively on the original's maximum difficulty.
1Epoch says it plans to release more than 1,000 gameplay videos and extend the harness to more games.
1
Mechanism
The benchmark couples model reasoning to screen interpretation and computer-use actions over a sequence of turns. Performance therefore reflects the complete harnessed system and its ability to sustain decisions, rather than an isolated answer or game-knowledge test.
Why now
Epoch's six-month comparison indicates that this task has shifted from near-inability to high success at the base setting, making evaluation of more difficult settings and system behavior newly consequential.
Implications
A visual game benchmark measures an agent system, not just the model, because success depends on perception, action selection, and persistence across a long task.
The next useful step is to compare traces and failure modes, not treat a single win rate as a general measure of autonomy.
Evidence boundary
The findings are specific to the tested games, selected models, difficulty settings, attempt protocol, and Epoch's harness; they do not establish general computer-use competence or explain the causal source of the gains.
Watch conditions
Whether Epoch's planned gameplay videos expose repeatable failure modes and whether the harness extension produces comparable results on other games.