Deep Dive · XiaoHu Explains

Anthropic's Official Playbook: Claude Fable 5 as "Advisor," Claude Sonnet 5 Does the Work and Eats Most of the Bill

In advisor mode, Fable 5 only weighs in; in orchestrator mode, Fable 5 delegates tasks — either way, the cheaper Sonnet 5 handles most of the load
Quick take
  • Anthropic's official developer account @ClaudeDevs shared two multi-model collaboration patterns the team uses internally, on July 7, 2026.
  • Advisor mode: the executor model Claude Sonnet 5 calls the advisor model Claude Fable 5 only at key decision points. On SWE-bench Pro, this combo hit about 92% of Fable 5's standalone score at about 63% of its cost, with Fable 5 called an average of roughly once per task.
  • Orchestrator mode: Claude Fable 5 acts as coordinator, breaking down tasks and dispatching them to Claude Sonnet 5 worker sub-agents running in parallel. On BrowseComp, this combo hit about 96% of Fable 5's standalone score at about 46% of its cost.
  • Both patterns run on Anthropic's official Claude Managed Agents (multi-agent sessions) capability: advisor mode maps to a specific beta API tool (the advisor tool), orchestrator mode maps to the multi-agent coordination feature — both are officially documented with sample code, not something the team hacked together on their own.
  • Sub-agents and advisors each get their own persistent session thread — going back to the same sub-agent for a follow-up doesn't require re-feeding the entire history and paying for it in full again.
⚑
The original post is frontline experience shared by Anthropic's own developer-facing account; the SWE-bench Pro / BrowseComp numbers are Anthropic's own benchmarks, not independent third-party evaluations. This piece adds the corresponding official API docs and an official open-source notebook example on top of the original post, filling in the mechanics, code samples, and real pitfalls for both patterns.
1Source

Who's giving this advice

Anthropic's official developer-facing account @ClaudeDevs shared two multi-model collaboration patterns the team uses internally, on July 7, 2026.

Split the work between two Claude models at different price points: the stronger Claude Fable 5 only chimes in or delegates at key moments, while the cheaper Claude Sonnet 5 carries the task from start to finish — so most of the bill lands on the cheaper side.
⚑
The source is Anthropic's official developer-facing account @ClaudeDevs, sharing the team's internal, everyday practice. The original post only gave two concrete benchmark numbers (SWE-bench Pro, BrowseComp) — the mechanics need the corresponding official docs to fill in, which the next two sections do.
Advisor Claude Fable 5 Just gives advice Occasional check-in (thin) Executor Claude Sonnet 5 Heavy output generation (thick)
Signature diagram: the advisor is only consulted a handful of times (thin arrows), while the executor model keeps churning out content (thick lines, flowing tokens).
2Mechanism 1 · Advisor Mode

How the advisor model and the worker model split the job

Advisor mode isn't something the team cobbled together — it's a formal beta feature in the Anthropic API, the advisor tool. The executor model Claude Sonnet 5 takes the task and works through it on its own, only checking in with the advisor model Claude Fable 5 once it hits a point that needs a judgment call, then picks the task back up with that guidance.

Core mechanism

A two-model advisor/executor split: the more capable Claude Fable 5 only "gives advice," never "writes content." All the token-heavy generation work goes to the cheaper Claude Sonnet 5, so the bulk of the bill lands at the executor model's price point.

This isn't something you hand-roll by firing off two separate requests from your own code: you declare an `advisor_20260301`-type tool in `tools`, specify which model the advisor uses, and Anthropic's server side handles the rest. The executor model decides when to check in just like calling any other tool; the server spins up a separate inference pass for the advisor model, which sees the executor's full conversation so far (system prompt, tool results, content already written), returns its judgment, and that gets dropped straight back into the executor's conversation for it to continue writing with. All of this happens within a single API request: