四款前沿模型设计评审横评：Astra 排第二

jinjingliang @JinjingLiang原推发布于 2026.10.07

在 X 上打开原推

I gave 4 frontier models the exact same design review tasks—here’s how they ranked: 1. GPT-6.1-sol xhigh (Codex) 2. GPT-6-astra high (Codex) 3. Opus xhigh (Claude Code) 4. Fable 5.1 high (Claude Code) How it works: I built an LLM Council using @orca_build's built-in orchestrator. A Sonnet 5 President agent triggers council members to run their own reviews, has them rate each other's work (blind), and aggregates the final votes. I’ve seen this consistently across multiple council runs—GPT models just do a better job with research & design tasks.