Claude Code tiers its models, with a nod to Jev for routing
Sonnet writes code, Haiku searches files in parallel and Opus reviews only three critical checkpoints; the post ends by noting the same idea applies to simple routing decisions in agent frameworks like Jev.
Most people are using Claude Code wrong. They're burning Opus-level intelligence on tasks Haiku can handle in parallel while Sonnet does the actual building. Here's a smarter setup: → Sonnet handles the main session, writes code and runs tests → Haiku spawns parallel subagents to search files, parse code and fetch docs → Opus acts as your architect, reviewing plans, debugging dead ends and catching regressions The trick? Opus doesn't need to be involved in every decision. Call it in at 3 critical moments: 1. Before locking the plan. Catch broken assumptions and architectural mistakes. 2. After the same test fails twice. Find the root cause instead of looping. 3. Before shipping. Review the full diff for hidden regressions. The result: faster execution, better parallelism, and your most capable model focused on decisions that actually need it. Sonnet builds. Haiku swarms. Opus reviews. The same principle applies to agent frameworks like JEV, where simple routing decisions can be handled without waking up a frontier model. You can set this up in Claude Code by asking it to configure the model roles and advisor checkpoints, then review every config diff before applying changes. Docs: https://code.claude.com/docs/en/advisor https://x.com/skunzu/status/2108265811421155637/video/1
This post quotes another post: x.com/arindam_1729/status/2108226317825302863