Introduction
Google published a model-routing study on November 3, comparing Gemini configurations with TypeSafe’s Jev classifier. The new material is an engineering experiment. Google’s study
What the experiment found
Both Flash with low thinking and a Jev-routed Flash-Lite/Flash setup answered all 50 test questions correctly. Routing reduced cost and waiting time against that baseline. Capping output length instead could truncate difficult answers without saving money. Results and configurations
That makes the comparison useful: test a lighter configuration before adding another service to the request path.
Google’s current API documentation exposes a thinking-level control for supported models, including Gemini 3.8 Flash. It is a separate setting from the maximum output length. Use the controls documented for the specific model and API rather than assuming a shorter response means less internal work. Thinking controls
Where a business could try it
Consider software that reads incoming service requests. Extracting an order reference and deciding how to resolve a disputed charge are different jobs. We would first try deterministic code for a reference with a known format. A model may help with an ambiguous message, while account changes still follow the application’s permission and review process.
For a routing trial, we would keep a small set of representative requests and compare:
- One model on the existing configuration.
- The same model with a lighter reasoning setting.
- A routing step followed by the selected model.
Count the entire attempt: classification, generation, retries and verification. Record wrong routes and incomplete answers beside the bill. Include requests that look short but require several checks, plus long requests that only need one field extracted.
This is how we would evaluate a proposed integration at araNET; it is not a claim that we have deployed this routing system or measured customer savings.
Availability and limits
Google’s companion workshop documents hosted Jev access through TypeSafe and Gemini access through Google Cloud. It also includes a rehearsal option without a live decision model. Live model use requires the relevant account access and can incur charges; a workshop example is not a ready-made connection to a business application. Official workshop
The study reports one run per configuration on 50 prompts with automatic grading. Its benchmark code awaits open-source review; the experiment date is unspecified. These results do not establish a savings rate for another workload. Study limitations
Repeated context also matters. Google documents cache-hit requirements and exposes cached-token usage in API responses. We would include that usage in the comparison, especially when requests repeatedly carry the same project information. Context caching
The decision to keep a router should come from verified results on the actual task. If a simpler configuration meets the same acceptance checks, the extra integration needs to earn its place.
