Gemini Decisions: the question has to fit the box

I’m interested in a model that chooses from a defined set without writing an essay. Gemini Decisions makes that a concrete API pattern; the first thing I’d test is whether it knows when my choices are inadequate.

What changed

Google’s 11 November announcement introduces the Gemini Decisions API, powered by Gemini Decision 1. Google says it can score up to 1,000 independent questions against shared multimodal input, with yes/no, multiple-choice, and ordered-rubric formats. It returns probabilities and confidence scores rather than generating a prose answer.

Where I’d put it

For a desk companion, I would consider a narrow routing task: does this request belong in the desk view, the markets view, or the voice workflow? That is a proposed use, not an integration I have built. The option set would also need a way to request clarification. A confident choice is not useful if none of the choices fits.

The test I’d actually run

I would prepare a small set of synthetic requests with intended routes, including ambiguous messages and requests outside the device’s scope. I would record wrong confident routes separately from uncertain ones, then test a threshold that sends uncertain requests back for clarification.

I would also compare the full interaction with a simple rule-based router. The question is whether the extra model call earns its place. No API test, latency measurement, or device result is claimed in this note.

Official source, published 11 November 2026: Gemini Decisions — read the original announcement/research ↗. Source checked 12 November 2026.