Jev plus Opus 5.5 layering: cheap model handles forks, expensive one only takes the long path
The author uses jev-opus to hand fork decisions like effort, tools, retries and escalation to Jev, leaving Opus 5.5 to write and plan, with per-message statements preserving the prompt cache.
It blows my mind how few people have wired this stack. A week of Jev plus Claude Opus 5.5 is enough to pull the cheap forks off the frontier model and let the expensive one think only when it has to. TypeSafe shipped Jev as a System One model that cannot write a sentence. Typed choices, scores, and probabilities in under 500ms at $0.042 per million input tokens. Anthropic shipped Opus 5.5 eight days later. Fable 5.1 level work, 40% cheaper than Opus 5, built for long agent runs. The logic is straightforward: You sit Jev on every fork (effort, tool, retry, escalate) and keep Opus 5.5 on the long path. Effort changes go in as per-message statements, so the prompt cache keeps growing instead of resetting. The system analyzes the turn, maps the next call, and compounds. More data in, smarter crew out. How it breaks down in practice: Install jev-opus and point Claude Code at claude-opus-5-5 Drop the template: raw for incoming files, wiki for compiled pages, CLAUDE.md for Opus, JEV.md for the typed questions Jev scores the prompt before the first token and again after every tool batch Opus writes, plans, and stays on the cached conversation Move a source into raw and issue one command. Query the crew instead of a blank canvas A five minute install means you stop burning Opus tokens on classification. Full walkthrough is in the article. Bookmark it.
This post quotes another post: x.com/merccante/status/2105040815537971408