Claude turns the advisor pattern into a native tool on Claude Platform

What Claude introduced

On April 9, 2026, Claude announced that the advisor strategy is now available in beta on the Claude Platform. The pattern is simple but strategically important: run Sonnet or Haiku as the executor that carries out the task, and let Opus step in only when the smaller model needs help deciding what to do next. Anthropic says this brings near Opus-level intelligence to agentic systems without paying Opus-level cost for the entire run.

The productization detail matters. Rather than asking developers to hand-roll a multi-model orchestration layer, Anthropic has turned the pattern into a native server-side tool called advisor_20260301. In Anthropic’s description, the executor keeps control of tools, reads results, and produces the user-facing output, while the advisor only returns a plan, correction, or stop signal. That means teams can adopt the pattern as a one-line change inside a single Messages API request instead of building a separate planner-worker stack.

What the early numbers say

Anthropic’s blog and follow-up post attach concrete evaluation data to the launch. The headline claim is that Sonnet with an Opus advisor scored 2.7 percentage points higher on SWE-bench Multilingual than Sonnet alone while costing 11.9% less per task. Anthropic also says Sonnet-plus-advisor improved results on BrowseComp and Terminal-Bench 2.0 while still coming in below Sonnet solo cost on a per-task basis.

The company argues that the same structure also changes the economics of small-model agents. In its published benchmark examples, Haiku with an Opus advisor reached 41.2% on BrowseComp versus 19.7% for Haiku alone. Anthropic frames that as a way to scale intelligence only when complexity demands it, rather than running an expensive frontier model end to end on every request.

Why this matters for agent architecture

The important shift is architectural. A lot of agent infrastructure over the last year has assumed that the smartest model should sit at the top as the orchestrator and delegate work downward. Anthropic is explicitly inverting that pattern. The cheaper model remains in the driver’s seat and escalates only when it hits a hard decision. That changes both cost control and system design.

• Developers can cap help requests with max_uses and monitor advisor tokens separately.
• The handoff happens inside one API request rather than across multiple round-trips and custom context plumbing.
• Frontier reasoning is applied only at the moments where it appears to matter most.

An inference from this launch is that model vendors are starting to compete not only on raw benchmark leadership but on intelligence routing: how gracefully they let teams mix models, tools, and budgets inside a single production workflow. If that trend continues, the most valuable platform feature may not be “use the biggest model everywhere,” but “use the right model at the right moment with the least orchestration overhead.”

Sources: Claude X post · Claude eval follow-up · Claude blog post