AI plays Slay the Spire

Results

Score by model and effort

Each dot is one run: hover it for details, click it to open the run. Shades are reasoning effort, from light (min) to full colour (max).

Ascension 20

GPT-6 Astra at max effort on the hardest difficulty, next to its Ascension 0 runs. Click a dot to open the run.

Score against cost

Mean score per model and effort against mean inference cost per run. Hover a line to pick out a model.

How the score works

Every run gets the score Slay the Spire itself gives at the end of a run, and the charts show that number unchanged. The models got the same formula in their instructions.

During the run

  • 5 per floor reached
  • 2 per normal enemy killed
  • 10, 20 or 30 per elite killed in Act 1, 2 or 3
  • 50 for the first boss, 100 for the second, 150 for the third and 200 for the fourth. Act 4 is enabled in every run.
  • 25 per elite beaten without taking damage
  • 50 per boss beaten without taking damage, or a flat 200 once three bosses are
  • 25 once for dealing 99 or more damage in a single hit
  • 25 once for playing 20 or more cards in one turn

Ascension

Everything above gets 5% more per Ascension level, so Ascension 1 adds 5% and Ascension 20 doubles it. The bonuses below are not scaled.

End of run

  • 25 for each non-starter card with 4 or more copies in the deck
  • 50 for holding 25 or more relics
  • 50 for a deck of 50 or more cards, or 25 for 35 or more
  • 50 for gaining 30 or more max HP over the run, or 25 for 15 or more
  • 100 for 5 or more curses in the deck
  • 75, 50 or 25 for earning 3,000, 2,000 or 1,000 gold over the run
  • 25 for visiting 15 or more "?" rooms
  • −1 for holding the Spirit Poop relic

Only on a win

  • 100 for a deck with no duplicate non-starter cards
  • 50 for a deck with no rare cards
  • 50 if the run took 45 minutes or less, or 25 for 60 minutes or less. The clock includes the model's thinking time.
  • 250 for beating the Corrupt Heart

Runs that end early

A run that dies, hits the turn or time limit, or stops on its own keeps the score it had at that moment, so climbing further always counts.