Google tests Jev routing for lower-cost Gemini coding agents
Google Developers Blog reports a 50-task, auto-graded coding-agent benchmark using live Gemini and Jev APIs to route tasks toward cheaper model configurations. Routing simple tasks to Flash-Lite and harder tasks to Flash with low thinking reportedly improved cost and response time, while limiting output tokens failed seven hard tasks. The results came from single-run tests on a sample suite, and one capped-output routing setup was slower and slightly more expensive than using normal Flash for every task.
The benchmark suggests that coding-agent developers should optimize cost per successful task and directly lower thinking levels before adding routing, while validating any classifier against their own workloads.
Cost per task: Using cheap classifiers to route across Gemini models and thinking levels - Google Developers Blog
Cost per task: Using cheap classifiers to route across Gemini models and thinking levels
NOV. 3, 2026
Sergio Villani Technical Solutions Google Cloud AI

#Google#TypeSafe#Jev#Google Antigravity#Gemini 3.8 Flash#Gemini 3.5 Flash-Lite#DiffusionGemma#Google Cloud