Google tests cheap classifier routing Gemini coding tasks across models and thinking levels
Overview
Google Developers Blog reports that a lightweight classifier can route coding-agent tasks between Gemini models and thinking levels.
In a test of 50 tasks, the router sent simple tasks to Flash-Lite and escalated hard ones to Gemini 3.8 Flash on low thinking, solving all 50, the same result as Flash on low thinking alone.
The post says a token output cap does not reduce thinking on easy tasks, so it argues cost per successful task matters more than price per token. The report comes from Google's own developer blog and covers a single test set of 50 tasks.
Written by AI from the articles below · updated Nov 3, 1:29 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- Google Developers BlogOfficialGoogle tests a cheap classifier for routing coding tasks across Gemini models
AIGoogle Developers Blog tested a lightweight classifier, Jev from TypeSafe, to route 50 coding-agent tasks between Gemini models. Jev solved all 50 tasks by sending simple ones to Flash-Lite and escalating hard ones to Gemini 3.8 Flash on low thinking, matching the 50 of 50 result of Flash on low thinking alone. The post says a token output cap does not reduce thinking on easy tasks, so cost per successful task matters more than price per token.
Heat Trend
Not enough continuous observations to show a trend yet.