Google Cloud reported that Jev routing and Gemini 3.8 Flash with low thinking each solved all 50 coding-agent tasks in its benchmark. The routed setup also had lower total billed cost and average latency in that run. Sergio Villani described the test in a Google Cloud report published Nov. 3, 2026.
How Jev routed coding tasks
The tested route sent each task to Gemini 3.5 Flash-Lite first. Jev, a structured-output classifier from TypeSafe, estimated whether a task needed more than one sequential calculation, simulation step, or logical deduction. Tasks it flagged as complex went to Gemini 3.8 Flash with low thinking.
Google Cloud reports that Jev matched hand-assigned difficulty labels on all 34 held-out tasks and 48 of the 50 tasks overall. Those figures describe the classifier’s agreement with the labels; the models’ task-solving results are a separate measure.
Both configurations solved 50 of 50 tasks
Here are the two setups that completed every task in Google Cloud’s 50-task run. The costs are totals for that test, and latency is the reported average per task.
| Configuration | Tasks solved | Total billed cost for the 50-task run | Average latency per task |
| Jev routing: Gemini 3.5 Flash-Lite first, then Gemini 3.8 Flash with low thinking for flagged tasks | 50/50 | $0.02287 | 1.90 seconds |
| Gemini 3.8 Flash with low thinking throughout | 50/50 | $0.02632 | 3.55 seconds |
The test showed a difference in reported cost and speed, not task completion: both configurations solved all 50. Google Cloud says each configuration ran once, using 50 prompts across seven task categories, with difficulty assigned by hand and simple automatic grading. The result is specific to that benchmark.
Context caching can change the tradeoff
Switching models can complicate long, context-heavy agent turns. Google says cached context can cost 75% to 90% less to resend, but each cache is tied to one model. A route that switches between Flash-Lite and Flash may therefore need separate caches or lose the discount.
