前 Anthropic 工程师泄露 Opus 5.5 对 GPT-6 策略对比文件
爆料称 Anthropic 缩模型、OpenAI 砍价格，两场发布押的是「天花板」还是「地板」，并给出 Sol 在编码与知识工作上退步的数据。
AN EX-ANTHROPIC ENGINEER LEAKED A FILE ABOUT COMPARING OPUS 5,5 WITH GPT-6 STRATEGY He was fired after that 16:31 Opus 5.5 → 18:12 Sol and Luna → 2 bets → 1 market → back to the price sheet Anthropic cut the model: Opus 5.5 is smaller than Opus 5 and still took the index at 58 OpenAI cut the price: $4 and $20 became $2 and $10, and the index moved 47 to 48 one bet says the ceiling sells, the other says the floor sells on 47 business tools the floor wins: 33.2% at $0.27 a task against 26.9% at 11× that on computer use nobody wins: 60.5% against 60.3%, and only the bill is different on coding the floor slipped: Sol scores 68.8% where its own last model scored 72.7% on knowledge work it lost about 100 Elo, which is what cutting the price actually cost 2 launches → 2 bets → 1 afternoon → 0 agreement on what a model is for the file does not say who wins, it says which bet you are already paying for read only the index and you will back the ceiling on work that never needed it I just continued using my Grok Bot agents, and it's still work the best
这条是引用着另一条写的：x.com/razeden0/status/2094391863897624907