Google launches Gemini Decisions API for fast, calibrated multimodal decisions

Google is carving decisions out of the general-purpose model stack. In a post on its AI Blog, the company introduced the Gemini Decisions API, describing it as powered by a model called Gemini Decision and built for fast, calibrated multimodal decisions.

The three adjectives in the launch framing carry most of the information. Fast points at latency, calibrated points at how well the model's stated confidence matches its actual accuracy, and multimodal points at inputs beyond plain text.

Calibration matters because it is what makes automation thresholds defensible. A well-calibrated model that says it is ninety percent sure should be right about ninety percent of the time, which is the property teams need before letting a model auto-approve, auto-route or auto-escalate a request.

The product also implies a separation of concerns. Instead of sending every short judgment through a large chat model, developers can call a dedicated endpoint and reserve generative models for long-form work, matching each job to a tool built for it.

That split is where the practical appeal sits. A specialised decision model can be tuned for latency and cost in ways a general model cannot, and it can be evaluated against a single, narrow task rather than a broad capability suite, which makes regressions easier to catch.

What the blog does not yet say matters as much as what it does. There are no published benchmarks, no pricing, no regional availability and no rate limits in the framing, so the near-term questions are whether Google will release calibration figures, how the endpoint is priced, and where the boundary sits between this API and the existing Gemini developer products.