Opus 5.5 decides while Jev executes in a split-role demo
A video demo has Opus 5.5 only read the inbox, set priorities and approve, while Jev writes code, drafts and runs research, using the split to cut output-token cost from the expensive model.
this is f*cking insane:1 Anthropic model thinks. 1 agent does the work. The expensive one never types. I made a video of the split. Orange is Opus 5.5. Purple is JEV. Opus 5.5 thinks: > reads the inbox > sets the priority > says go or no go JEV does: > writes the code > writes the drafts > runs the research Opus 5.5 price per million tokens: $4 in, $20 out. An output token costs 5x more than an input token. Read the fine print. The numbers on screen are a demo of the pattern. Your real cost depends on your prompts. Paste this rule into your agent: Opus 5.5 plans and decides only. Send code, drafts and research to the doer. Come back to Opus only for the final check. Most people let their smartest model do the typing
This post quotes another post: x.com/thegreatest_sv/status/2102802378378621355