Run Claude Code on Sonnet and keep Opus on call as an advisor
You don't need the expensive model on every turn. With Claude Code's advisor tool, a Sonnet main session does the work and consults Opus only before it commits to a plan, when an error keeps coming back, and before it says it's done.
The problem
Most of a coding-agent session is routine: reading files, small edits, running tests. Running all of it on the strongest model burns quota or budget on turns that don’t need it. Running everything on a cheaper model saves money until it confidently picks the wrong plan.
The tip
Split the roles. @Voxyz_ai’s post, with close to 1,000 bookmarks, shows the setup: put the main session on Sonnet 5.5 and attach Opus as an advisor. Sonnet keeps writing the code. At decision points it calls Opus, which reads the full conversation, including tool calls and results, and returns guidance.
Per Anthropic’s docs, Claude tends to consult the advisor before committing to an approach, when an error keeps recurring, and before declaring a task done. Advisor calls bill at the advisor’s rates, but only happen at those moments. Toggling /advisor doesn’t invalidate the main prompt cache, which switching models with /model does.
Example
# one session
claude --model sonnet --advisor opus
# or make it the default in ~/.claude/settings.json
{
"model": "sonnet",
"advisorModel": "opus"
}
Then add one line to CLAUDE.md so the timing is explicit:
Consult the advisor before committing to a large plan, when the same error appears twice, and before calling a long task done.
Things to check: the advisor must rank at or above the main model (Haiku 4.5 can’t be an advisor). It needs the Anthropic API or Claude Platform on AWS, not Amazon Bedrock, Google Cloud’s Agent Platform or Microsoft Foundry. Setting CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1, or DISABLE_TELEMETRY (which blocks the feature-flag fetch), turns it off silently.
When NOT to use it
Skip it for short tasks with little to plan, since the advisor adds tokens without much to decide. Skip it too when every turn needs deep reasoning; switch the main model instead.