Anthropic’s Advisor: multi‑model orchestration
Anthropic’s Advisor approach calls a heavy reasoning model and lighter supporting models in a single API call, treating model tiering as a runtime pattern rather than just procurement. The write‑up explains pairing Opus 4.6 with cheaper models (Sonnet or Haiku) inside one request and explicitly ties the pattern to regulatory considerations like the EU AI Act. That design highlights platform needs: traceability of which model handled each subtask, cost attribution, and standardised orchestration primitives. (decodethefuture.org)
Anthropic put a two-model workflow into one Claude API call on April 9, letting a cheaper model do most of the work and ask Opus for help only on harder decisions. (claude.com) Large language models generate text token by token; Anthropic’s new “advisor tool” lets one model act like the day-to-day worker while a stronger model steps in as a strategist mid-task. Anthropic says the pattern is in public beta and is aimed at long-running agent workloads such as coding agents, computer use, and multi-step research pipelines. (platform.claude.com) In Anthropic’s setup, Claude Sonnet 4.6 or Claude Haiku 4.5 serves as the executor, and Claude Opus 4.6 serves as the advisor. The advisor reads the full conversation and returns a plan or course correction, typically 400 to 700 text tokens, before the executor continues. (platform.claude.com) Anthropic says developers enable the feature with the beta header `advisor-tool-2026-03-01` and add a server-side tool with the type `advisor_20260301`. The company’s release notes list the launch date as April 9, 2026, and describe the goal as getting close to advisor-solo quality while keeping most generation at executor-model rates. (platform.claude.com) The pricing gap is what makes the pattern notable. Anthropic lists Claude Opus 4.6 at $5 per million input tokens and $25 per million output tokens, while Claude Sonnet 4.6 starts at $3 per million input tokens and $15 per million output tokens. (anthropic.com 1) (anthropic.com 2) Anthropic frames the tool as a way to avoid an all-or-nothing choice between a premium model and a cheaper one. Its product page says developers can pair Opus with Sonnet or Haiku and get “near Opus-level intelligence” at a lower cost for agents. (claude.com) That is a shift in how model tiers are sold to developers. Instead of buying one model for one endpoint, Anthropic is packaging tiering as runtime behavior inside the request itself, with the platform deciding when the stronger model is consulted. (platform.claude.com 1) (platform.claude.com 2) The compliance angle is not hypothetical in Europe. The European Union’s Artificial Intelligence Act includes transparency duties for some systems under Article 50, and a separate public tracker lists that article’s date of entry into force as August 2, 2026. (ai-act-service-desk.ec.europa.eu) (artificialintelligenceact.eu) A multi-model request makes basic record-keeping more concrete: which model handled which subtask, which tokens were billed at which rate, and what advice changed the final output. Anthropic’s own tool documentation already distinguishes between server tools Anthropic executes and client tools an application executes, which is the kind of boundary auditors and enterprise buyers usually ask to see. (platform.claude.com 1) (platform.claude.com 2) Anthropic has spent the past year pushing Claude toward agents that plan, use tools, and run longer jobs, from expanded agent documentation to a 1 million token context window in beta for Opus 4.6 and Sonnet 4.6. The advisor tool fits that push by turning model mixing from custom glue code into a product feature. (anthropic.com) (anthropic.com) (anthropic.com) The immediate test is whether developers trust Anthropic’s routing enough to build around it. If they do, “use Opus or use Sonnet” stops being a procurement choice and starts looking more like traffic management inside the API. (claude.com) (platform.claude.com)
