testingcatalog.com
|
ksl
|
|
Anthropic released an advisor tool in its API that lets developers pair Opus as a reasoning layer with Sonnet or Haiku as the executor – escalating to the more expensive model only when the cheaper one hits a decision it can’t handle alone. The benchmark numbers are striking: Haiku with Opus advising more than doubled its standalone score on SWE-bench Multilingual while costing less than running Sonnet by itself. It’s a routing architecture baked into the Messages API rather than something developers need to build from scratch. The pattern of tiered model orchestration is spreading fast – OpenAI ships similar logic inside Frontier, and Google has been doing model cascading in Vertex for months – but packaging it as a first-party primitive lowers the bar for teams building cost-sensitive agents.
