{"rewrite":{"id":"r_e73e8c2714006b860d9799f0","clusterId":"c_ba71e21cf256273756289f60","slug":"ai-models-clear-slay-the-spire-s-first-act-then-stall-at-ascension-20","model":"deepseek-v4-1-flash","headline":"AI Models Clear Slay the Spire's First Act, Then Stall At Ascension 20","summary":"A preprint paper on arXiv put 20 large language models through Slay the Spire at Ascension 20, the game's highest difficulty, using the Ironclad and the same 10 seeds. GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5 each cleared Act 1 in 9 of 10 runs. Only GPT-5.6 Sol beat the final boss, once.","whyItMatters":"The study's own framing is that state-of-the-art models top the average human but fall far short of skilled players, and a model that plays safe enough to never lose early may never be willing to take the risks the later acts demand.","webCardHtml":"\u003cp\u003eThe models were run through SpireBench, an environment built by extending an existing game operation mod so an LLM can play Slay the Spire directly. Because each model may already carry strategy knowledge for the game, the main experiments replaced in-game names with different ones.\u003c/p\u003e\u003cp\u003eResults were set against roughly 206,000 Ironclad play records released by Mega Crit and the play histories of four veteran streamers. The authors note that 10 trials per model produce large variance from luck, so the paper is not reliable for ranking models that finish close together.\u003c/p\u003e","blueskyPost":"A preprint put 20 LLMs through Slay the Spire at Ascension 20. Three cleared Act 1 in 9 of 10 runs. Across 200 total attempts, the final boss died once, to GPT-5.6 Sol. The authors say 10 trials is too few to rank models that finish close together.","twitterPost":"20 LLMs played Slay the Spire at Ascension 20, 10 runs each. GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5 cleared Act 1 in 9 of 10 runs. Across all 200 attempts the final boss fell once, to GPT-5.6 Sol. The paper calls 10 trials too few to rank them.","threadsPost":null,"newsletterBlurb":"A preprint on arXiv had 20 large language models play Slay the Spire at Ascension 20, ten runs each with the Ironclad. The best models cleared the first act almost every time and then stalled: only GPT-5.6 Sol beat the final boss, in a single run. The authors compare the results to roughly 206,000 human Ironclad records from Mega Crit and to four veteran streamers, and caution that ten trials is too small a sample to rank models that land close together.","attributionJson":"[{\"source\":\"Automaton\",\"url\":\"https://automaton-media.com/articles/newsjp/20261014-473406/\",\"title\":\"Study Has Latest AI Models Play 'Slay the Spire' at Highest Difficulty, Finds They 'Struggle a Lot in the Later Half'; Too Safety-First, So They Die\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":4714,"outputTokens":673,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1791963431,"createdAt":"2026-10-14T07:31:47.000Z","publishedAt":"2026-10-14T07:36:06.000Z","updatedAt":"2026-10-14T07:36:06.000Z"},"cluster":{"id":"c_ba71e21cf256273756289f60","canonicalTitle":"『Slay the Spire』最高難易度を最新AIモデルにプレイさせると“後半でやたら苦戦する”との研究。安全第一すぎて死ぬ","representativeArticleId":"a_ae5d184a37bfe3b885008270","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[\"Slay the Spire\"],\"studios\":[],\"people\":[],\"type\":\"news\",\"domain\":\"games\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-10-14T06:15:25.000Z","lastSeenAt":"2026-10-14T06:15:25.000Z","updatedAt":"2026-10-14T07:36:06.000Z"},"attribution":[{"source":"Automaton","url":"https://automaton-media.com/articles/newsjp/20261014-473406/","title":"『Slay the Spire』最高難易度を最新AIモデルにプレイさせると“後半でやたら苦戦する”との研究。安全第一すぎて死ぬ"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":["Slay the Spire"],"studios":[],"people":[],"type":"news","domain":"games","is_roundup":false},"keyFacts":null}