SpireClimb (Slay the Spire 1 and 2) (SpireClimb)
We show this table for reference; we do not rank on it.
Epoch AI's SpireClimb has models play Slay the Spire and Slay the Spire 2 at each game's maximum difficulty through a computer-use harness, as an end-to-end test of computer use, long-horizon reasoning and tactical play.
Benchmark score on SpireClimb — November 13, 2026
We compile the SpireClimb rows from benchmark-owner or independent runs. GPT-6 Astra leads the table at 34.0%, followed by GPT-6.1 Sol (23.8%) and GPT-6.1 Terra (22.3%). We do not use these results to rank models overall.
GPT-6 Astra
OpenAI
Run by SpireClimbGPT-6.1 Sol
OpenAI
Run by SpireClimbGPT-6.1 Terra
OpenAI
Run by SpireClimb9 modelsAgenticCurrentDisplay onlyUpdated November 13, 2026
Benchmark score table (9 models)
| Rank | Model / configuration | Score | Who ran it | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | GPT-6 AstraOpenAIRun by SpireClimb | 34.0% | Run by SpireClimb | Not reported | Closed |
| 2 | GPT-6.1 SolOpenAIRun by SpireClimb | 23.8% | Run by SpireClimb | Not reported | Closed |
| 3 | GPT-6.1 TerraOpenAIRun by SpireClimb | 22.3% | Run by SpireClimb | Not reported | Closed |
| 4 | Claude Fable 5.2AnthropicRun by SpireClimb | 19.6% | Run by SpireClimb | Not reported | Closed |
| 5 | GPT-6 SolOpenAIRun by SpireClimb | 9.2% | Run by SpireClimb | Not reported | Closed |
| 6 | Claude Opus 5.5AnthropicRun by SpireClimb | 5.8% | Run by SpireClimb | Not reported | Closed |
| 7 | Claude Fable 5.1AnthropicRun by SpireClimb | 2.0% | Run by SpireClimb | Not reported | Closed |
| 8 | Claude Sonnet 5.5AnthropicRun by SpireClimb | 1.3% | Run by SpireClimb | Not reported | Closed |
| 9 | GPT-6 LunaOpenAIRun by SpireClimb | 0.0% | Run by SpireClimb | Not reported | Closed |
- Lab-reported
- Benchmark-run
- Independent
How to read this leaderboard
Higher is better. A score is a share of won runs, so it is a strict pass/fail measure of finishing a whole game, not of progress within one. GPT-6 Astra leads with 34.0% (33.3% in StS1, 34.7% in StS2), still far below top human players. BenchLM stores each game's win rate in the row notes.
Operator receipt: 9 sourced rows are currently displayable on this page; the leading published row is GPT-6 Astra at 34.0%.
Honest limit: Nine models so far, each at one effort setting with one harness. Both games are popular, and Epoch notes labs may train on them, so the benchmark could saturate quickly. The frozen Slay the Spire 2 build will drift from the version people play. BenchLM shows SpireClimb as display-only; it does not enter overall or category rankings.
Among the reported SpireClimb rows, GPT-6 Astra is first at 34.0%. The third row is 11.7 points behind. The broader top-10 range is 34.0 points, so the table still separates the published systems.
9 models have been evaluated on SpireClimb. The benchmark falls in the Agentic category. We keep these results separate from the weighted global scoring system, so they remain source-specific evidence. SpireClimb is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About SpireClimb
Year
2026
Tasks
60 Slay the Spire 1 runs (Ascension 20, Heart kill) and 75 Slay the Spire 2 runs (Ascension 10) per model
Format
Win rate, averaged across the two games, computer-use harness
Difficulty
Maximum game difficulty, long-horizon computer use
Models play through a computer-use harness. In Slay the Spire 1 the setting is Ascension 20 with the Act 4 Heart kill: 15 runs on each of the four characters, 60 runs in total. In Slay the Spire 2 it is Ascension 10: 15 runs on each of five characters, 75 runs. Every model plays the same fixed seeds. The SpireClimb score is the average of the two win rates, with both games weighted equally. Epoch froze Slay the Spire 2 at version 0.107.1 so the task stays fixed while the early-access game keeps changing. Top human players win roughly 85% of max-difficulty runs in the first game and about 95% in the second.
Freshness and provenance
Version
SpireClimb 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does SpireClimb measure?
Whether a model can win full games of Slay the Spire and Slay the Spire 2 at maximum difficulty while operating the game through a computer-use harness. It combines computer use, long-horizon planning and tactical card play.
Which model leads SpireClimb?
GPT-6 Astra at max effort leads with 34.0%, ahead of GPT-6.1 Sol (23.8%), GPT-6.1 Terra (22.3%) and Claude Fable 5.2 (19.6%). Top human players win roughly 85% to 95% of runs at these settings.
How is the score calculated?
Epoch averages two win rates: Slay the Spire 1 at Ascension 20 including the Heart (60 runs) and Slay the Spire 2 at Ascension 10 (75 runs). Both games count equally.
Compare top models on SpireClimb
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.