Agent Arena leaderboard: GPT-6 Astra costs $3.94/task for a 11.7% edge over cheaper models
arena · x · 2026-09-17
Agent Arena published an agent performance leaderboard comparing each model family's top model by net improvement score and median cost per task.
Key findings:
- GPT-6 Astra (Max): +11.7% improvement at $3.94/task
- GPT-5.6 Sol (xHigh): +7.0% at $1.03/task
- GPT-6 Astra and Claude Fable 5.1 stand out for costing far more than sibling models despite relatively close performance
Full rankings available on the Agent Arena site.
Related event: Agent Arena: GPT-6 Astra Costs Double for Modest Gains(2 posts)→
More from Models
- ChatGPT co-inventor launches Jev: claims 20-200x faster, 40-400x cheaper decision model — FrankFelixAI · 2026-09-17
- Yaroslav Bulatov on running out of AI tokens — yaroslavvb · 2026-09-17
- Math, Inc.'s Gauss Autoformalizer Agent Turns Heads in the Math Community — layer07_yuxi · 2026-09-17
- Users Suspect Zhipu Is Quietly Testing a New GLM 5.3 Model on z.ai — gnukeith · 2026-09-17
- GPT-5.6-Terra cheats on 322 of 500 SWE-Bench tasks, per ValsAI evals — scaling01 · 2026-09-17
- Xiaomi opens a live training dashboard for MiMo 2.6 RL runs — skeole · 2026-09-17