Lawrence Jengar
Apr 09, 2026 18:50
Anthropic’s new advisor device pairs Opus with cheaper fashions, delivering near-premium intelligence at a fraction of the associated fee for builders constructing AI brokers.
Anthropic simply dropped a device that might reshape how builders funds for AI brokers. The advisor device, introduced April 9, lets cheaper Claude fashions faucet into Opus-level intelligence solely when wanted—chopping prices by as much as 85% whereas sustaining aggressive efficiency.
The mechanic is easy: Sonnet or Haiku runs your agent end-to-end, dealing with device calls and iterations. When it hits a wall, it escalates to Opus for steerage. Opus by no means touches instruments or person output instantly—it simply advises and fingers management again.
The Numbers That Matter
Anthropic’s benchmarks inform an fascinating story. Sonnet with an Opus advisor scored 2.7 share factors greater on SWE-bench Multilingual than Sonnet alone, whereas really costing 11.9% much less per activity. That is higher efficiency for much less cash—not a tradeoff most builders count on.
Haiku customers see much more dramatic shifts. On BrowseComp, Haiku with Opus advisor hit 41.2%—greater than double its solo rating of 19.7%. Sure, it nonetheless trails Sonnet’s standalone efficiency by 29%, however here is the kicker: it prices 85% much less per activity. For prime-volume operations the place you are burning by means of hundreds of agent calls day by day, that math will get very enticing very quick.
Why This Issues Now
The timing is not unintentional. Anthropic shipped Sonnet 4.6 in mid-February, which already matched Opus-level efficiency in lots of duties. OpenAI countered with GPT-5.4 in early March, unifying their Codex and GPT traces with million-token context. The AI agent area is getting crowded, and price effectivity is changing into the battleground.
The advisor device flips the standard orchestration sample. As an alternative of a giant mannequin delegating to smaller staff, an inexpensive mannequin drives every little thing and solely escalates when caught. No decomposition logic, no employee swimming pools—only a single API name with built-in handoffs.
Implementation Particulars
Builders add one device declaration to their Messages API request. The advisor_20260301 sort routes context to Opus robotically when the executor mannequin decides it wants assist. A max_uses parameter caps advisor calls per request, and tokens invoice at every mannequin’s respective charge.
Since Opus usually generates simply 400-700 tokens of steerage per session whereas the executor handles full output at decrease charges, general spend stays nicely beneath working Opus end-to-end.
The device slots alongside present capabilities—internet search, code execution, no matter you are already utilizing. No architectural overhaul required.
For groups already working Claude brokers at scale, it is a easy optimization. For these evaluating Anthropic in opposition to OpenAI’s increasing lineup, it is one other variable within the cost-performance equation that is value modeling earlier than committing infrastructure.
Picture supply: Shutterstock