Splitting Claude Code across opusplan and Haiku subagents (the Explore subagent needs its own switch)

This setup plans on Opus 5.5, writes code on Sonnet 5.5 and runs subagents on Haiku 5.5.

A post on X billed as a setup "Anthropic just documented" says to start with this line.

CLAUDE_CODE_SUBAGENT_MODEL=haiku claude --model opusplan

opusplan is a real, documented alias: it uses opus in plan mode and sonnet when executing. Three of the post's claims don't match the docs.

• Repository search doesn't move to Haiku. On a Claude subscription or Console account, the built-in Explore subagent runs on Opus, and CLAUDE_CODE_SUBAGENT_MODEL alone doesn't change the Explore or Plan subagents.
• Opus doesn't step in "when the same test fails twice" or "before signing off on the diff". opusplan switches only on whether you're in plan mode. To have Claude consult Opus mid-task, turn on the advisor (/advisor opus).
• fallbackModel doesn't guarantee availability. It switches only when a model is overloaded or unavailable, never on rate-limit, authentication or billing errors.

To put every subagent on Haiku, use your settings file (v2.1.257 or later; Haiku 5.5 needs v2.1.293 or later).

// ~/.claude/settings.json
{
  "model": "opusplan",
  "env": {
    "CLAUDE_CODE_SUBAGENT_MODEL": "haiku",
    "CLAUDE_CODE_SUBAGENT_MODEL_FORCE": "1"
  },
  "fallbackModel": ["sonnet", "haiku"]
}

Without CLAUDE_CODE_SUBAGENT_MODEL_FORCE, Explore, Plan and any subagent whose definition sets its own model keep their models. To move only exploration to Haiku, define your own Explore subagent with model: haiku instead.

Haiku 5.5 runs with a 1M-token window by default, but prompts over 100K tokens cost more. Open /tasks to see which model each subagent is running on.