Anthropic Launches Advisor Tool for Claude Platform API
Summary
- • Anthropic releases advisor tool letting Opus guide Sonnet or Haiku on-demand via the Claude Platform API
- • Executor models handle routine tasks and escalate to Opus only at complex decision points
- • Haiku with Opus advisor more than doubled benchmark scores while costing less than running Sonnet alone
- • Enabled with a simple Messages API configuration change requiring no re-architecture
Details
Advisor tool publicly available on Claude Platform API
Activated via a simple configuration in the Messages API request — no significant rearchitecting needed for existing Claude Platform users. The executor and advisor share full task context, allowing Opus to review the complete state before returning a plan or corrective guidance.
Small executor models escalate to Opus only at complex decision points
This inverts traditional orchestration where large models delegate down to smaller ones. Sonnet or Haiku runs autonomously and pulls in Opus only when it hits a decision point beyond its capability — keeping most per-task cost at the cheaper executor level.
Haiku + Opus advisor more than doubled standalone benchmark score at lower cost than Sonnet
Improvements measured on SWE-bench Multilingual (coding), BrowseComp (web navigation), and Terminal-Bench 2.0 (terminal tasks). The augmented Haiku setup cost significantly less than running Sonnet alone, making it compelling for cost-sensitive production deployments.
Targets developers building cost-sensitive AI agents at scale
Primary use cases: automated coding assistants, research pipelines, multi-step workflow automation where reasoning quality matters but end-to-end frontier model use is prohibitively expensive. Anthropic positions this as infrastructure for scalable, cost-efficient agent deployment.
Product Launch = feature/tool release; Tech Info = architectural details; Stat = benchmark and cost results; Strategy = positioning and target use cases
What This Means
Developers building AI agents can now get near-Opus reasoning quality at closer to Haiku prices by letting a cheaper model do most of the work and call in the powerful model only when it gets stuck — a practical cost-performance tradeoff that makes advanced AI agents more accessible at scale.