AI Got Crushed by Slay the Spire’s Hardest Mode? I Looked at the Data, and the Headline Falls Apart

200 runs. One win.

That is the full record of 20 large language models against the final boss on Slay the Spire’s A20 difficulty. GPT-6 Astra, Claude Fable 5.1, and GPT-5.6 Sol trade places atop other leaderboards. Against the Corrupt Heart, their combined win rate was 0.5%.

News headlines say AI got crushed. Break the data apart, and that sentence is only one-third right.

1. 90%, 10%, 0.5%, 4%

Radboud University and the Vector Institute ran the study. The paper is on arXiv. The design was simple: 20 models, 10 identical random seeds each, all playing the Ironclad.

Top models cleared Act 1 at 90%. That rate is close to veteran players. Act 2 separates them. By Act 3, GPT-6 Astra drops to 10%. Claude Fable 5.1 sits at 20%. Against the final boss, GPT-5.6 Sol won once in 200 runs.

If final boss means the Act 3 boss, a 10% to 20% clear rate is not ugly. If it means the Corrupt Heart, a 0.5% heart-kill rate is low.

Mega Crit gave the researchers 2020 data. Among all players attempting A20, Act 1 clear rate was 47%, Act 2 was 15%, Act 3 was 7%, and heart kill was 4%. Among players with 1000+ hours, the numbers are 54%, 20%, 11%, and 7%.

Apart from heart-kill rate, top AI models already beat the average 1000-hour human player on Act 3 clear rate.

2. A20 clear rate 7.3%, heart-kill rate lower

The Slay the Spire community has a saying: A20 is where the game starts.

Steam achievement data tells another story. Players who clear A20 make up 7.3% of game owners. If players who clear A0 count as effective players, the A20 clear rate is 10.9%. If the “Kill the Guardian” achievement is the effective-player baseline, that number drops to 8.5%.

These statistics only count cleared A20. They do not count consistent heart kills. Clearing the Act 3 boss on A20 and defeating the Corrupt Heart are separated by key collection, route planning, and deck ceiling.

A20 NOSL win rates above 20% are expert-circle records, not per-capita reality. Before using that standard to mock AI, check whether you are in that 7.3%.

3. 68%, 19.3, 23%, 16

The data shows where AI loses.

At rest sites, 68% of AI choices are rest. Human players choose rest 23% of the time. AI challenges elites far less often than humans. Average deck size is 19.3 cards for AI and 16 for humans.

These three numbers point to one behavioral pattern. AI chooses short-term safety at every local decision point. It rests to heal. It skips elites. It takes a card whenever offered. Each step looks lower-risk. Together, they produce a bloated deck, a lower draw rate for key cards, too few relics, and not enough upgrade slots.

By Act 3 and the Heart, deck strength is insufficient. Precise calculation cannot fix that.

The failure sits in cross-act planning. A 90% Act 1 clear rate shows AI can calculate damage, allocate energy, and choose targets. It cannot invest in the self that will exist twenty floors later.

4. 10 seeds, no S/L, context limits

The study has three boundaries.

First, each model ran 10 seeds. In a 200-run sample, the difference between 0.5% and 1% is not meaningful. Run 30 or 50 more, and the numbers change.

Second, AI did not use S/L. In Slay the Spire 1, S/L can change card generation and target selection. Human experts use it to optimize decisions repeatedly. AI played one-life runs. Human data is not necessarily all NOSL.

Third, the models did not undergo extensive Spire training. Context length also limits their ability to remember decisions from dozens of floors earlier. Give them tutorial cases, more training, and search, and results would differ.

The paper authors’ conclusion is narrower than the headlines:

Frontier models beat rule-based and symbolic baselines and the average human player, yet remain far from expert play.

5. Tactical giant, strategic dwarf

AI’s performance in Slay the Spire comes down to one phrase: tactical giant, strategic dwarf.

It knows the optimal card to play this turn. It does not know what cards to take this run, which elites to fight, or which rest site to use for an upgrade instead of sleep. Its decision horizon covers the current fight. It stops there.

Rest 68% vs 23%, deck 19.3 vs 16. These two gaps matter more than the Act 3 clear rate. Act 3 clear rate is affected by seeds and luck. Rest rate and deck size are stable outputs of model strategy.

The paper’s claim is narrower: AI beat average players. It did not beat experts.

If that counts as crushed, then human players with a 4% A20 heart-kill rate, including me, have been crushed too.