ROASBench: Score

Top 10 of 30 models · score out of 100, higher is better · ≈ cannot be told apart from the leader

  1. 1≈ Claude Fable 5.2Anthropic 61.3
  2. 2≈ GPT-6.1 TerraOpenAI 58.4
  3. 3≈ GPT-6 AstraOpenAI 56.6
  4. 4 Kimi K3.1 ProMoonshot AI 55.1
  5. 5 GPT-6 SolOpenAI 54.1
  6. 6 GPT-6 LunaOpenAI 53.4
  7. 7 Claude Sonnet 5.5Anthropic 52.0
  8. 8 Muse EmberMeta 51.8
  9. 9 Claude Opus 5.5Anthropic 51.7
  10. 10 Claude Fable 5.1Anthropic 51.5

Results to 9 Nov 2026 · release 8342a7178929Source: Spring Prompt · CC BY 4.0