Frontier models now win at Slay the Spire

In roughly six months, frontier language models moved from rarely completing a run of Slay the Spire to winning over 90% of games at the base difficulty. At the hardest setting, GPT-6.1 Sol now wins one run in five. Epoch AI reports the results in SpireClimb, a benchmark for long-horizon planning and visual computer use.

Slay the Spire is a single-player roguelike deck-builder in which each run presents a changing sequence of cards, enemies, rewards, and hidden future events. Decisions made early can determine whether a deck survives much later encounters. Previous reinforcement-learning systems and language-model agents struggled with that combination of randomness, incomplete information, and interdependent choices, making the game a useful complement to short-form chat benchmarks.

Six months to a 98% win rate

In June 2025, the Orak paper reported that Gemini 2.5 Pro, operating through a text harness, defeated the second boss in two of three runs. The paper did not report a completed-run victory. An Anthropic announcement earlier in 2026 said Claude Fable 5 reached the final act three times as often as Claude Opus 4.8, which still fell short of reliable wins.

Epoch’s latest results show a different performance range. At Ascension 0, the base game without cumulative difficulty modifiers, leading frontier models win at least 90% of their runs. GPT-6 Astra leads with a reported 98% win rate, followed by GPT-6.1 Sol and Claude Opus 5.5.

Inside the pixel-and-mouse test

Epoch’s maximum-difficulty protocol uses a single-agent visual computer-use harness based on Inspect ReAct. The model receives screenshots, reasons about the visible state, and operates the game through mouse clicks rather than reading structured game data from an API.

Each model receives 15 independent attempts per character on a fixed set of seeds selected in advance. That produces 60 attempts across the original game’s four characters and 75 across the sequel’s five characters.

• Visual input: The model sees the same graphical interface as a human player.
• Mouse control: It selects cards, routes, rewards, and other actions by clicking the interface.
• Persistent notes: It may record information that survives context compaction, the process of compressing earlier interactions when the context window fills.
• Reference access: It may consult the Slay the Spire wiki, giving it access to information available to an experienced player.
• Single-shot runs: Each attempt starts from the beginning, with no reload from a saved state after a mistake.

Epoch designed the setup to approximate what an unskilled user might obtain by asking a desktop agent to play the game autonomously. The score therefore reflects strategic reasoning, visual perception, interface control, memory management, and effective use of external references.

Ascension 20 separates the field

Ascension 20 applies twenty cumulative difficulty modifiers, including stronger enemies and harsher constraints. The “A20 Heart” evaluation also requires taking the optional route to The Heart and defeating it after the normal final act.

Strong human players regularly report A20 Heart win rates above 50%. XecnaR, widely regarded as one of the best players, has reported rates from 80% to 95% depending on the character. Those human results come from broader play rather than Epoch’s fixed seeds, so they provide context rather than a controlled comparison.

Against that standard, GPT-6.1 Sol won 20% of its A20 Heart runs and GPT-6 Astra won 17%. Every Claude model tested recorded zero wins at this setting.

Sol’s character-level results broadly followed human experience. It won 33% of Watcher runs and 27% of Ironclad runs, while Silent produced no wins in 15 attempts. Epoch does not yet have a clear explanation for the Silent result.

Fresh characters test transfer

Because the original game is nearly a decade old, its strategies, card rankings, and encounter guides may appear extensively in model training data. Epoch included the early-access sequel to reduce that overlap and test whether models could transfer their skills to newer content.

Frontier models exceeded a 25% win rate at Slay the Spire 2’s current maximum difficulty, Ascension 10. GPT-6 Astra reached 35%. The sequel remains easier overall in its early-access form, so its percentages are not directly comparable with A20 Heart results from the original game.

Astra won 53% of its runs with Regent, one of the sequel’s new characters, and 13% with the new Necrobinder. Its results for the three returning characters fell between those figures. Strong performance with Regent weakens a pure-memorization explanation, although the benchmark cannot isolate novel reasoning from transferred strategy, wiki use, or public material published after the sequel’s release.

The signal beyond the game

SpireClimb tracks several capabilities relevant to autonomous software agents:

• maintaining a goal across long sequences of dependent decisions;
• planning under uncertainty and recovering from unfavorable outcomes;
• extracting useful state from a changing visual interface;
• controlling software through ordinary mouse interactions;
• preserving information through external notes and context compaction; and
• adapting established strategies to unfamiliar characters and mechanics.

Desktop agents need a similar combination of perception, planning, memory, and interface control to complete multi-step workflows without a custom API. Product teams still need domain-specific evaluations for reliability, security, and error recovery because a game score does not establish performance on business software.

Where the score can mislead

• Sampling noise: The original game provides 60 maximum-difficulty attempts per model. One additional win changes the aggregate rate by about 1.7 percentage points, so the gap between 17% and 20% represents only a few runs.
• Entangled capabilities: The graphical harness makes perception and mouse control part of the score, preventing a clean attribution of gains to strategic reasoning alone.
• Open-book scaffolding: Wiki access and persistent notes make the benchmark a test of tool-assisted agency rather than unaided recall.
• Fixed-test exposure: Published seeds support consistent comparisons and create a target for optimization once laboratories train directly against the benchmark.
• Saturation risk: Explicit training on Slay the Spire could push base-difficulty results to the ceiling within one or two model generations.

Epoch plans to evaluate new models as they appear, publish more than 1,000 recorded gameplay videos, and extend the harness to additional games. The recordings will allow researchers to inspect individual decisions and distinguish planning failures from perception, control, and memory errors.