Anthropic launched the Advisor Strategy on April 9, 2026 — a new architecture pattern available on the Claude Platform API that lets developers get near-Opus intelligence in their AI agents without paying Opus prices for every token. The idea is simple but the benchmark results are significant: pair Claude Opus 4.6 as a strategic advisor with Claude Sonnet 4.6 or Haiku 4.5 as the executor, and the cheaper model escalates to Opus only when it genuinely needs help.
What Is the Advisor Strategy?
The Advisor Strategy is a new API pattern where two Claude models work together in a single API call that shares full context. The executor model — Sonnet 4.6 or Haiku 4.5 — handles the bulk of tasks independently: managing tools, running routine operations, generating outputs. When the executor encounters a decision it cannot handle confidently, it escalates to Opus 4.6 as the advisor. Opus reviews the full shared context, returns a plan or corrective feedback, and the executor resumes the task with that guidance. The key distinction from traditional orchestration is direction: instead of a large model delegating down to smaller agents, a smaller model runs independently and escalates up only when necessary. This keeps most operations at executor-tier pricing while accessing Opus-level reasoning for the moments that actually require it.
How to Enable It
The advisor tool is available publicly through the Claude Platform API and can be activated with a simple configuration in the Messages API request. Developers add the advisor tool to their existing API setup — no separate endpoint, no architectural overhaul required. The full documentation is available at platform.claude.com/docs/en/agents-and-tools/tool-use/advisor-tool.
Benchmark Results
Anthropic released benchmark data comparing standalone models against the advisor pattern across three evaluations.
| Setup | SWE-bench Multilingual | Cost Change |
|---|
| Sonnet 4.6 alone | 72.1% | Baseline |
| Sonnet 4.6 + Opus advisor | 74.8% (+2.7pts) | -11.9% per task |
| Haiku 4.5 alone | Low baseline | Cheapest |
| Haiku 4.5 + Opus advisor | More than doubled | Less than Sonnet alone |
The Haiku result is the most striking — Haiku with an Opus advisor more than doubled its standalone benchmark score while costing significantly less than running Sonnet alone. This makes the advisor pattern particularly compelling for high-volume workloads where Haiku is already the default choice for cost reasons.
Why This Matters for Developers
The cost structure of Claude API pricing makes the advisor strategy financially significant. Claude Opus 4.6 costs $5 per million input tokens and $25 per million output tokens. Claude Sonnet 4.6 costs $3 input and $15 output. Claude Haiku 4.5 costs $1 input and $5 output. Running Opus for every token in an agentic workflow is expensive at scale. The advisor pattern means you pay Haiku or Sonnet rates for the majority of operations and only incur Opus costs for the subset of decisions that genuinely require flagship reasoning. The 11.9% cost reduction on SWE-bench tasks — which are computationally intensive coding evaluations — suggests the savings will be meaningful across most production agent workloads.
How It Differs From Traditional Orchestration
Most multi-model AI architectures use a large model as the orchestrator — it plans, delegates subtasks to smaller models, and synthesizes results. The advisor pattern inverts this. The smaller, cheaper model is always in control and always executing. Opus is a resource called on demand, not a manager directing from above. This inversion has practical advantages: the executor maintains full task context, escalations are targeted rather than broad, and the system spends the minimum time at Opus-level pricing. Anthropic describes this as letting the executor model escalate only when necessary, keeping most operations at a lower cost.
Real-World Use Cases
The advisor pattern is well suited to any agentic workflow where most steps are routine but occasional steps require genuine reasoning depth. Software development agents — the SWE-bench results confirm this is the primary validated use case. Haiku handles file reads, simple edits, and routine code generation. Opus steps in for architecture decisions, complex debugging, and multi-file refactors. Customer support automation — Haiku handles standard queries and lookup tasks, escalating to Opus for edge cases requiring nuanced judgment. Research and analysis agents — Sonnet handles data gathering and summarization, Opus handles synthesis and strategic conclusions. Document processing pipelines — Haiku processes high-volume routine documents, Opus handles exceptions and complex cases.
Pricing Context
With prompt caching, the advisor strategy becomes even more cost-effective. Anthropic offers 90% savings on cached prompt tokens — repeated context shared between executor and advisor calls costs significantly less on subsequent requests. The Batch API adds an additional 50% discount for non-time-sensitive workloads. Combined with the advisor pattern, developers can reduce agentic workflow costs by 60-80% compared to running Opus for all operations while maintaining near-Opus output quality on benchmark tasks.
What Developers Are Saying
Early reaction from the developer community has been positive. The consensus from practitioners is that the advisor pattern formalizes what many teams were already doing manually — routing difficult decisions to larger models — but makes it a first-class API feature with shared context and clean escalation logic. One widely shared observation: most Claude Code sessions could benefit from this pattern immediately — Haiku for high-volume, low-stakes operations, Sonnet for the bulk of real development work, and Opus reserved for architecture and serious decisions.
Availability
The advisor tool is available now through the Claude Platform API. It works with Claude Opus 4.6 as the advisor and either Claude Sonnet 4.6 or Claude Haiku 4.5 as the executor. The feature is accessible to all API customers with no waitlist or special approval required.