GPT-6 Astra 在 BrokenArXiv/ArXivMath 登顶，成本仅对手九分之一

Lumina @LuminaBench原推发布于 2026.09.16

在 X 上打开原推

🚨 GPT-6 Astra is #1 on the new BrokenArXiv / ArXivMath benchmark The cost difference is ridiculous: • Astra scores 81.9% while costing just $2.26 • Fable 5.1 scores 79.8% for $19.57 I find it absolutely crazy that Astra scores higher while costing almost 9x less mainly due to token efficiency

这条是引用着另一条写的：「BrokenArXiv 与 ArXivMath 新基准：Astra 居首」