The development

What happened

Google has introduced the Gemini Decisions API and Gemini Decision 1, an AI model that scores supplied options instead of generating an open-ended answer. The endpoint supports routing, classification and rubric-based evaluation over shared multimodal input. Google’s speed and calibration claims are developer-reported, not guarantees for every workload.

Key facts

  • The API accepts text, images, audio, video and PDFs, with up to 256K input tokens and 1,000 independent questions per request. Those are API limits, not the configuration used for every published performance result.

  • Its three question formats estimate a yes/no probability, choose among 2–255 candidates, or return a weighted score over an ordered rubric. Results include candidate probabilities and confidence values.

  • Google reports 150–250ms median end-to-end latency for tests with contexts up to 16K tokens, remaining nearly flat from one to 100 questions. That tested range is narrower than the maximum supported input and question counts.

  • At announcement, Google lists $0.10 per million input tokens and no output-token charge. Shared input is billed once per request; each question still adds instruction and criteria tokens. A no-cost prototyping tier is offered.

  • The model is available through the Gemini API in Google AI Studio. Ecosystem integrations are rolling out; Google Cloud support is described as coming soon. Google reports calibration results, but this article does not independently validate them.

The context

Why it matters

A bounded decision interface can separate an AI system’s routing or grading step from free-form generation. Returning probabilities also makes fallback rules possible, but a confidence value is not itself proof of reliability on a team’s own data. Deployment decisions still need task-specific error, calibration and latency checks.