Code4Sceneby SimWorld
Gallery

Every scene, every agent

What each configuration built on the 20 Research Track scenes in the October 2026 Text-to-Scene re-run, as the evaluator saw it: four renders per scene from the benchmark's own cameras, with the case score and the agent's cost. Scores are truncated to three decimals; a case without a valid scene scores 0.

In leaderboard order. The strip shows the same five scenes for every agent: Blue City, Cyberpunk, Egyptian Temple, Nordic Harbour, Winter Town.

  1. 1Claude Sonnet 5.5 (max)AnthropicScore0.837Median cost$26.93
  2. 2Claude Opus 5.5 (max)AnthropicScore0.832Median cost$26.90
  3. 3Claude Fable 5.1 (max)AnthropicScore0.800Median cost$23.42
  4. 4GPT-6 Astra (max)OpenAIScore0.790Median cost$14.25
  5. 5GPT-6.1 Sol (max)OpenAIScore0.784Median cost$2.59
  6. 6Gemini 4 Argon (high)GoogleScore0.777Median cost$4.83
  7. 7GPT-5.6 Sol (max)OpenAIScore0.738Median cost$6.87
  8. 8Claude Haiku 5.5 (max)AnthropicScore0.718Median cost$2.43
  9. 9Muse Wonder Full 1.0 (max)MetaScore0.715Median cost$3.66*
  10. 10Grok 4.7 (xhigh)xAIScore0.700Median cost$3.98
  11. 11Grok 4.6 (xhigh)xAIScore0.696Median cost$7.41
  12. 12Muse Spark 1.3 (max)MetaScore0.696Median cost$3.24*
  13. 13Gemini 3.8 Flash (high)GoogleScore0.694Median cost$3.85
  14. 14GPT-6 Luna (max)OpenAIScore0.688Median cost$0.16
  15. 15GLM-5.3 Flash (max)Z.aiScore0.663Median cost$0.94
  16. 16MiMo V2.6 Pro (max)XiaomiScore0.632Median cost$0.92
  17. 17Muse Spark Open (max)MetaScore0.611Median cost$2.01
  18. 18Gemma 4 31B (thinking off)GoogleScore0.553Median cost$0.04*
  19. 19Nemotron 3.5 Super VL (max)NVIDIAScore0.520Median cost$0.14
  20. 20Gemma 4 31B (thinking on)GoogleScore0.517Median cost$0.06*
  21. 21DeepSeek V4.1 Flash (max)DeepSeekScore0.371Median cost$0.35