Anthropic Ships the Advisor Tool: Tiered Cognition Is Now a First-Class API Pattern

Every conversation I've had about AI costs in the last six months eventually lands on the same tension: you want the smartest model for the hard decisions, but you can't afford to run it on every token. Anthropic just gave that tension a formal solution.

The advisor tool, now in beta under anthropic-beta: advisor-tool-2026-03-01, lets you pair Opus as a strategic advisor with Sonnet or Haiku as the task executor. The executor runs end-to-end on its own, and only consults Opus when it hits a decision point it can't resolve. Think of it like a junior engineer who handles the implementation but walks over to the senior's desk when the architecture gets weird.

The benchmarks tell the story. Sonnet paired with an Opus advisor gained +2.7 percentage points on SWE-bench Multilingual while cutting cost per agentic task by 11.9%. That's better AND cheaper simultaneously, which almost never happens. Haiku's results are even more dramatic: with Opus advising, it scored 41.2% on BrowseComp, more than double its solo 19.7%. Haiku went from "can barely browse the web" to "competent web researcher" just by knowing when to ask for help.

I've been building something like this manually in my own pipelines for months, using cheaper models for routine research and escalating to Opus for synthesis and judgment calls. Having it as a native API primitive changes things. You don't have to build the routing logic yourself anymore. The model decides when it's stuck.

The bigger pattern here is what I'd call tiered cognition architecture. It mirrors how every effective engineering team actually works. You don't have your principal engineer reviewing every line of code. You have them review the design decisions, the ambiguous requirements, the things where judgment matters more than execution speed.

If you're building agents today, here's what to do: audit your current Opus usage and identify which calls are genuinely hard decisions versus routine execution. Swap routine calls to Sonnet or Haiku with the advisor tool enabled. You'll likely see 10-15% cost reduction with equal or better quality on the tasks that matter.

The uncomfortable question this raises: if the advisor pattern works this well, what does that say about how much of our "hard" work is actually routine?