Gemini Decisions API turns multimodal judgment into an API primitive

Generative AI does not always need to produce a paragraph of text. In model routing, policy enforcement, tool selection, and content grading, an application often needs only a probability, a selected option, or a score. Google’s Gemini Decisions API is designed for those bounded decisions rather than open-ended generation.

What the API provides

• Many questions over one input: A request can attach up to 1,000 independent questions to the same shared context. This can avoid repeatedly submitting identical multimodal material.
• Native multimodal evaluation: The input can contain text, images, audio, video, and PDFs, including combinations such as a call recording with account metadata or a scanned claim with damage photos.
• Three decision modes: A yes/no mode estimates whether a condition holds. The choice mode selects among two to 255 candidates and returns normalized probabilities for each. The score mode maps an input to an ordered rubric and returns a weighted score together with level probabilities. The source’s sample uses the field name “noul” for the yes/no type, so implementers should follow the current field definitions in the official documentation.
• Low-latency operation: Google reports 50–100 ms of median model compute and 150–250 ms end-to-end latency. The model accepts contexts of up to 256K tokens, while the latency description specifically discusses contexts up to 16K tokens.
• Input-oriented pricing: The announced price is $0.10 per one million input tokens, with no charge for output tokens. Shared input is billed once per request, and Google says a no-cost tier is available for prototyping.

Instead of decoding an answer token by token, Gemini Decision 1 scores candidate outcomes in a single forward pass. The resulting confidence values can be used to send uncertain cases to a stronger model, a rule system, or a human reviewer. That makes the API particularly relevant to agent pipelines where predictable response time matters as much as language quality.

Why it matters

If Google’s stated performance and pricing hold in production, the API could serve as a lightweight control plane for generative applications. Simple requests could be routed to a cheaper model, while difficult cases are escalated. Policy checks and content classification could also be separated from verbose answer generation. Google cites internal examples including model routing in Gemini Notebook, rapid interaction editing in Stitch, emoji selection in CC, and safety checks for tools in Search.

There are important limits. A confidence score is not proof of correctness, and a closed set of candidates cannot represent every real-world situation. Teams will still need to test thresholds, calibration, error costs, and fallback behavior on their own data. In high-risk domains, the published benchmark and pricing claims should not replace independent evaluation.

Developers can access the API through Google AI Studio and use Python, JavaScript, or REST guides. Google also lists integrations with OpenRouter, LangChain, Vercel, LiteLLM, and other tools. The main proposition is not to replace a general-purpose model, but to extract frequent “which option?”, “does this pass?”, and “what level?” decisions into a faster, more economical layer.

Source: Google AI Blog