Skip to main content

The Effort Dial and Model Splitting

Spend frontier-model tokens on judgment and small-model tokens on commodity steps, and turn the effort dial down where deep thinking is wasted.

Hand-drawn line illustration in warm terracotta and cream: a loopy line-drawn hand turns a large plain cream dial with no markings while three cream blocks of increasing size sit in a row beside it, a one-line profile face watching.

Two levers, one idea​

Not every step in an agent's work deserves the same intelligence. The platform exposes that as two levers:

  • The effort dial. The Messages API takes an effort setting (levels from low up through the maximum) that trades thoroughness for speed and tokens; lower effort means fewer, more consolidated tool calls and less preamble. Per the platform docs (verify against platform.claude.com/docs; parameter names and levels drift), high effort suits agentic coding and hard reasoning, low effort suits subagents and simple steps.
  • Model splitting. Route each step to the cheapest model that clears its quality bar: a frontier model for judgment calls, a small fast model for extraction, classification, and formatting. Anthropic's post The advisor strategy (April 2026) formalizes one version: a more capable model advises and plans, a cheaper model executes.

The same logic gates whole architectures. Building multi-agent systems: when and how (January 2026) reports that "multi-agent implementations typically use 3-10x more tokens than single-agent approaches for equivalent tasks," which makes multi-agent a spend you justify, not a default.

The split in production​

  • Elation Health runs its clinical chart summaries on Claude Haiku 4.5, the small fast model, because in-visit latency is the constraint, and reports "Reduced median time to first insight by 61% for chart review." (source)
  • Eve Legal makes Opus its production default for legal judgment and "uses Claude Sonnet for lighter extraction work" across the 12.5 million documents it processes a month. (source)
  • Replit splits by difficulty inside one product: Sonnet handles sustained development work while Opus takes architectural decisions and complex multi-file refactoring. (source)
  • Asana states the principle outright: "By using different Claude models, we're able to optimize for speed, reasoning power, or balance between the two." (source)

How to find your split​

Work backward from an eval, not forward from a price sheet. Run each pipeline step against the smallest model at low effort, measure, and promote only the steps that fail. Teams that skip the measurement default everything to the frontier model, which is the expensive kind of caution. The eval harness that Outcomes and Strength of Evidence both argue for is the same harness that makes downgrading safe.

Further Reading​