20 个 Astra agent 对战 20 个 Claude agent
作者按 Terminal-Bench 4.0 上 57.9% 对 55.8% 的接近分数，做了场 20v20 脚本化 agent 对战，Astra 方 1-0 险胜。
OPENAI JUST DROPPED GPT-6 ASTRA, SO I SIMULATED 20 OF ITS AGENTS AGAINST 20 CLAUDE FABLE 5.1 AGENTS OpenAI reports 57.9% for Astra on Terminal-Bench 4.0, versus 55.8% for Fable 5.1. I turned that near-tie into a controlled 20v20 arena: one staff hit eliminates an agent, there are no respawns, and both teams are divided into four five-agent squads connected through a shared command bus. The agents coordinate targets, launch synchronized assaults, dodge incoming charges, rescue trapped teammates, and surround isolated opponents. Every command visible in the console changes the formations and tactics unfolding inside the arena. Fable takes the lead twice and forces a 3v3 endgame. Then the last GPT-6 agent gets trapped in a 2v1, parries the first strike, reverses the surround, and wins 1-0. The result is scripted, so this isn’t a model benchmark. But the coordination is the part worth watching.
这条是引用着另一条写的：x.com/lagerskoy/status/2094861375764152640