Bookmarked* (*not tested)

A viral Claude Code tip pairs an Opus advisor with Haiku subagents in one launch command. Half of it matches Anthropic’s docs. The other half we could not find there.

A dark book tied shut with cord, an asterisk on the cover, and a fringe of clay bookmark ribbons hanging below
Saved is not the same as checked.

A one-line setup tip is easy to save. It is harder to check, and most of us skip that part.

What happened

On October 8, Guillermo Casaus (@_guillecasaus) posted a Claude Code tip on X, in Spanish. The post says the idea is to combine Sonnet 5.5, Haiku 5.5 and Opus 5.5 “so that each model handles what it does best” (our translation), and tells readers to start Claude Code like this:

claude --advisor opus --subagents haiku

The post spread quickly. It is a reasonable idea, and the first half is real.

Anthropic documents the advisor tool, which “lets Claude consult a second, typically stronger model at key moments during a task.” The docs list three ways to set it, including “--advisor flag: set the advisor for a single session at launch,” with the example claude --advisor opus. The same page marks the feature as experimental and says it “requires the Anthropic API.”

The second half is harder to confirm. Anthropic’s subagent and model configuration pages describe two ways to choose a subagent’s model: a model field in the subagent’s definition, which accepts sonnet, opus, haiku, fable, a full model ID, or inherit, and the CLAUDE_CODE_SUBAGENT_MODEL environment variable. We found no --subagents flag on either page.

On October 10, 2026, we ran claude --help on version 2.1.296. The output lists no --subagents option and no --advisor option. That second absence is expected: Anthropic’s advisor page says Claude Code “doesn’t list --advisor in claude --help.”

The asterisks

1. Not in the help text is not the same as rejected. Because --advisor works without appearing in --help, a missing entry alone does not prove --subagents is fake. It does mean there is no documented promise behind it. We could not find one in the help output or on the docs pages we checked.

2. Our test could not catch a silent no-op. On version 2.1.296, claude --subagents haiku --help printed the normal help text and exited with code 0. claude --subagents haiku --version printed the version and exited 0. A made-up flag of our own did the same. These commands exit before a session starts, so they show nothing about what a real launch does with an unknown flag. We did not start a session, because that would mean logging in and making API calls.

3. The documented path has an order of precedence. Per Anthropic, Claude Code picks a subagent’s model in this order: a model Claude passes for that invocation, then the definition’s model field, then CLAUDE_CODE_SUBAGENT_MODEL, then the main conversation’s model. The docs add that CLAUDE_CODE_SUBAGENT_MODEL “is a default,” so a definition or a model Claude passes still wins. To force one model onto every subagent, the docs say to also set CLAUDE_CODE_SUBAGENT_MODEL_FORCE to 1, which requires v2.1.257 or later.

4. Version floors apply. Anthropic’s model configuration page says: “Use v2.1.293 or later with Haiku 5.5.” An older install can fail on the model the tip assumes.

What this means if you ship

A bookmarked one-liner tends to become trusted setup. It goes into a shell alias or a client’s onboarding doc, and nobody looks at it again.

If the subagent half does nothing, the cost of that is quiet. You believe your subagents run on Haiku, and you price your work on that belief. Meanwhile, by Anthropic’s documented order, a subagent with no other assignment runs on the main conversation’s model. Nothing in the tip tells you which one you got.

None of this is a knock on the person who shared it. The advisor half is documented, and routing routine work to a cheaper model is a sound goal. The point is narrower: a launch command is a configuration claim, and a configuration claim should be checked against the docs and against a running session before it touches a bill.

What to do Monday

  1. Check your version. Run claude --version. Anthropic says to use v2.1.293 or later with Haiku 5.5, and claude update upgrades.
  2. Read claude --help yourself, and remember that --advisor is documented even though it is not listed there.
  3. Set the subagent model the documented way. Add model: haiku to a subagent’s definition, or set CLAUDE_CODE_SUBAGENT_MODEL to haiku. Add CLAUDE_CODE_SUBAGENT_MODEL_FORCE set to 1 if you want it applied to every subagent.
  4. Confirm what actually ran. Anthropic’s docs say to run /tasks while a subagent is running: “The subagent’s row shows the model it runs on.” That requires v2.1.242 or later.
  5. Keep the advisor if you want it. Launch with --advisor opus, or set it with /advisor. The docs mark it experimental and say it requires the Anthropic API.

Bookmark the tip. Then run the check.

Sources

Next: GPT-6* (*which one?)