把 GPT-6 Astra 只用在最后 10% 的省钱架构
作者主张别让前沿模型互相比拼，而是让 Kimi K3 铺量、Astra 只做最终把关，示例成本从 180 美元降到约 37 美元。
THE SMARTEST AI STACK MAY SPEND 90% OF ITS TOKENS ON THE CHEAPER MODEL AND SAVE THE BEST MODEL FOR THE LAST 10%. Everyone keeps asking which frontier model wins. I think that's becoming the wrong question. Kimi K3 can fan work out across up to 300 agents. GPT-6 Astra is much better suited to one expensive problem where a wrong decision matters. So don't make them compete. Make them work different shifts. In one illustrative workload: Astra for everything: $180 K3 with repeated context cached: ~$32.40 Then put Astra only on the final gate: +$4.50 K3 researches, scans, drafts and throws away bad attempts. Astra gets the final deliverable and one instruction: “Find what's wrong.” That's the architecture I find much more interesting than another benchmark leaderboard. The most capable model doesn't need to do the most work. It should do the work where being wrong costs the most. Cheap intelligence for width. Expensive intelligence for judgment. ROUTE TOKENS LIKE YOU ROUTE CAPITAL.
这条是引用着另一条写的：x.com/slash1sol/status/2098013767627837827