Google's dev blog reports that cost per successful task, not per-token price, now dominates agent economics, and shows a lightweight classifier (TypeSafe's Jev) routing simple tasks to Flash-Lite and hard ones to Gemini 3.8 Flash with thinkingLevel "LOW" solved all 50 benchmark tasks while reducing cost and latency versus running low thinking everywhere. Key findings: capping maxOutputTokens doesn't reduce thinking spend and broke 7 of 25 hard tasks, keyword heuristics dropped to 70.6% on held-out tasks, and routing across models conflicts with Context Caching since caches are per-model. Deve…
Read the full story on developers.googleblog.com
What to doSet thinkingLevel: "LOW" on Flash before building any router, and if you add classifier-based routing, route only short standalone tasks while keeping long context-heavy turns on a single model to pr…
- Provider