Google Gemini Decisions API: Scored AI Decisions for 1,000 Questions

What happened

Google has announced the Gemini Decisions API, which lets developers have AI applications make routine decisions without generating a text response, and can evaluate up to 1,000 questions in a single request. The launch came weeks after OpenAI introduced a competing service, according to InfoWorld.

The API runs on Google's new Gemini Decision 1 model. Conventional generative models produce output token by token. This system instead assigns probabilities to a set of options the developer specifies, so an application can act on a result without waiting for a full generated answer.

Developers can use the API to check whether a condition has been met, to choose one option from a predefined list, or to score an input against a rubric that measures how well it meets set criteria. It accepts multimodal inputs, including images, audio and video, with a context window of up to 256,000 tokens. The resulting confidence scores can guide when an application should act on a decision.

Why it matters

Much of today's AI activity is spent producing prose when the underlying task is simply triage: is this claim complete, which queue should this request join, does this recording contain a complaint? A decision-oriented interface turns those judgements into structured, scored outputs that software can route on immediately. That shifts the cost and latency profile of AI from writing to deciding, which matters for any organisation running high volumes of customer or case interactions.

For leaders in experience, AI and digital transformation, the important question is less about the model and more about the operating model it enables. Scored decisions across text, images, audio and video could allow service workflows to be redesigned around thresholds and escalation rules, rather than around staff reading and interpreting AI-generated summaries.

By the numbers

• 1,000 questions can be evaluated in a single Gemini Decisions API request.
• 256,000 tokens is the maximum context window supported.

The Renascence take

Most coverage will frame this as a developer tool for faster, cheaper AI. The more useful reading is that it makes automated judgement a first-class design element, and that raises the stakes for how those judgements are set up.

The real risk is not that the model is wrong; it is that a confidence score looks like an answer, so teams stop questioning it. Customer-obsessed operators should treat every threshold as a service decision, with a named owner, a visible fallback for low-confidence cases and regular testing against what customers actually experienced. Behavioural economics tells us people anchor on the first number they are shown, so design what the customer sees, not just what the system computes. Build the decision logic around the moments that matter to the customer before you automate the rest.