Skip to content

Atomic Task Checks for Model and Thinking-Level Routing

Atomic task checks route work using a specific reasoning requirement, then evaluate savings against verified task outcomes and a fixed-model baseline.

An atomic routing check asks about one task property instead of requesting a broad difficulty judgment. Google's experiment replaced a simple-versus-complex classification with a check for sequential calculations, simulations, or deductions (Google report).

Use a competitive baseline

The experiment used 50 short coding-related questions with automatic grading. A frozen keyword-and-length heuristic matched difficulty labels on 70.6% of 34 held-out questions; Jev matched all 34. Label agreement measures routing decisions, not answer correctness (experiment).

Gemini 3.8 Flash at low thinking answered all 50 correctly without routing. The successful routed setup first tried Flash-Lite, then escalated tasks Jev classified as complex to Flash at low thinking. It also answered all 50 correctly. The study reports one run per setup and hand-assigned difficulty labels (results and limitations).

These results support testing a narrow classifier. They do not establish that a router preserves quality on repository-scale changes.

Count the entire attempt

Use total spend divided by verified successes. Include the first answer, classifier, escalation, and retries. A cheaper failed answer must remain in the numerator.

Lowering maxOutputTokens did not reduce thinking cost in the reported easy-task comparison, and the small cap broke hard tasks. The escalation target therefore needs its own measured configuration (configuration comparison).

For your own evaluation, freeze the routing question before the held-out test. Track unnecessary escalations separately from hard tasks sent to the cheap path. Compare both against the cheapest fixed configuration that meets your quality threshold.

Do not transfer the short-task result to context-heavy turns without measuring cache effects. Model switching can require another model's cache; the source explicitly limits its routing recommendation to short, self-contained tasks (cache discussion).

Key Takeaways

  • Choose the fixed-model baseline before calculating routing savings.
  • Report answer correctness separately from classifier agreement.
  • Include the first cheap attempt when computing escalation cost.
Emerging — single recent source Last reviewed: 2026-11-06