Back to explore

Jev小型モデルがCodexで推論強度を動的調整しAstra請求を削減

Somi氏の共有:Jev小型モデルがCodexでGPT-6と連携し、ステップごとに推論強度を調整。vechenのテストでAstra請求を50%削減し、プロンプトキャッシュも維持。

50% off the Astra bill in vechen's tests, prompt caching intact. Jev, a tiny model, sits beside GPT-6 in Codex and dials reasoning effort per step. Stuck? Think harder. Routine? Don't. GPT-6 is a bad judge of when it's stuck, and I'd rather a cheap watcher own that knob.

· 0 likesOpen on X