Advisor Mode
An experimental dual-model collaboration mechanism in Claude Code in which a primary model performs coding tasks while a stronger reasoning advisor model intervenes only at key moments, such as architectural decisions, complex debugging, and final review, to offer suggestions without writing code directly.
Timeline (last 90 days)
Anthropic introduced an experimental tool, Advisor Mode, for the Claude Code terminal, whose core idea is to let a strong reasoning model intervene only at key decision points rather than participating in coding throughout.
Advisor Mode is offered through the Anthropic API and the Claude Platform on AWS, but it is not yet supported on Amazon Bedrock, Google Cloud Agent Platform, and Microsoft Foundry.
Advisor Mode is billed at the token pricing of the model being invoked and is deducted from the account balance; it is not a free add-on feature.
There are three ways to enable Advisor Mode: type /advisor in a session to bring up the model list, type /advisor opus or /advisor fable to specify a model and save it to settings, edit settings.json to set the advisor model key, and use the claude --advisor opus startup flag to enable it temporarily.
The hard rule of Advisor Mode is that the advisor model's capability tier cannot be lower than the main model's; if an advisor weaker than the main model is forcibly specified, Claude Code simply disables the advisor feature rather than crashing.
Claude Code supports only one main model paired with one advisor model per session, and the "four-model pipeline" circulating in the community is not an official one-click mode.
Each time the advisor model is invoked, it re-reads the entire session history from scratch, and the advisor's own reads are not cached, while the main model's cache remains intact.
A community X post claims Advisor Mode can save 3.4 times the cost, but the official documentation does not give that figure and only confirms the directional conclusion that invoking an expensive model at key decision points can reduce costs.
When the advisor is invoked, the screen shows an advising line with the model name, and after it finishes it shows Reviewed, with Ctrl+O available to view the full advice; refused guidance shows declined, and an invocation error shows unavailable.
You can turn the advisor off within a session with /advisor off, or set the environment variable CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1 to remove the tool entirely at the system level.
1 more timeline events
All Facts (11)
The hard rule of Advisor Mode is that the advisor model's capability tier cannot be lower than the main model's; if an advisor weaker than the main model is forcibly specified, Claude Code simply disables the advisor feature rather than crashing.
60%UnverifiedWhen the advisor is invoked, the screen shows an advising line with the model name, and after it finishes it shows Reviewed, with Ctrl+O available to view the full advice; refused guidance shows declined, and an invocation error shows unavailable.
60%UnverifiedYou can turn the advisor off within a session with /advisor off, or set the environment variable CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1 to remove the tool entirely at the system level.
60%UnverifiedAnthropic introduced an experimental tool, Advisor Mode, for the Claude Code terminal, whose core idea is to let a strong reasoning model intervene only at key decision points rather than participating in coding throughout.
60%UnverifiedClaude Code supports only one main model paired with one advisor model per session, and the "four-model pipeline" circulating in the community is not an official one-click mode.
50%UnverifiedEach time the advisor model is invoked, it re-reads the entire session history from scratch, and the advisor's own reads are not cached, while the main model's cache remains intact.
50%UnverifiedA community X post claims Advisor Mode can save 3.4 times the cost, but the official documentation does not give that figure and only confirms the directional conclusion that invoking an expensive model at key decision points can reduce costs.
50%UnverifiedIn Advisor Mode, when the advisor is triggered frequently, the overhead of repeatedly re-reading the session history may offset most of the cost savings.
50%UnverifiedAdvisor Mode is offered through the Anthropic API and the Claude Platform on AWS, but it is not yet supported on Amazon Bedrock, Google Cloud Agent Platform, and Microsoft Foundry.
50%UnverifiedAdvisor Mode is billed at the token pricing of the model being invoked and is deducted from the account balance; it is not a free add-on feature.
50%UnverifiedThere are three ways to enable Advisor Mode: type /advisor in a session to bring up the model list, type /advisor opus or /advisor fable to specify a model and save it to settings, edit settings.json to set the advisor model key, and use the claude --advisor opus startup flag to enable it temporarily.
50%