Google's Gemini Decisions API Answers 1,000 Questions in One Call

Google on November 11, 2026 introduced the Gemini Decisions API, an endpoint powered by the new Gemini Decision 1 model that evaluates as many as 1,000 independent questions across text, images, audio, video and PDFs in a single request. The endpoint is live through the Gemini API in Google AI Studio.

The release targets a specific gap in AI application architecture. Developers building agents and routing systems routinely ask a language model a bounded question — which tool to call, whether a policy is violated, which result ranks first — and then parse generated prose back into structured values before the program can act. Gemini Decision 1 removes that round trip. Instead of decoding answer text token by token, the model scores candidate options in one forward pass, returning typed responses that programs can branch on, sort and route directly.

What can the Decisions API do?

The API handles bounded judgments: tool selection, policy checks, model routing, context grading and ranking. It does not generate prose or explanations. A request pairs one shared input — a document, an image, a video — with a set of independent questions evaluated against that input.

The Decisions API documentation sets a 262,144-token context window that covers both the shared input and all questions. Requests can mix three question types:

• Yes/no checks for binary validation.
• Fixed-set choices accepting two to 255 options.
• Ordered rubric scores, returned as probability-weighted averages of level indices with per-level probabilities attached.

Every answer carries a confidence score from 0 to 1, defined as the estimated probability that the top answer is correct. The design lets applications act automatically on high-confidence results and escalate uncertain ones to a human or a second check. Google's documentation recommends calibrating thresholds with labeled examples drawn from a developer's own traffic, and retaining human review for high-stakes actions such as payouts.

The questions are isolated from one another. Adding or removing a question does not change any other answer, which rules out dependent chains inside a single request. Dependent decisions belong in application logic: developers can send a second request after reading the first answer, or ask both questions together and discard the answer they do not need.

How fast is it, and what does it cost?

Google reports 50–100 milliseconds of median model-compute latency and 150–250 milliseconds of end-to-end latency across multimodal inputs and contexts up to 16K tokens. Latency stayed nearly flat between one and 100 questions per request — the marginal cost of an additional question is close to zero in time terms.

Gemini Decision 1 also led Google's internal benchmark suite spanning agent routing, reasoning, ranking and multimodal understanding, and posted the lowest calibration error of the models tested. That calibration figure matters directly for the confidence-score design: a model that misreports its own certainty would undermine the escalation pattern the API is built around.

Pricing is $0.10 per 1 million input tokens, with zero output tokens billed. The shared input is billed once per request regardless of question count; each question adds its own instruction and criteria tokens. There are no variable thinking-token charges and no separate audio-transcription or OCR costs. Google describes Gemini Decision 1 with the Decisions API as nearly five times cheaper across the shared benchmark suite than Gemini 3.5 Flash-Lite paired with the Interactions API — a comparison that positions the new endpoint against the cheapest existing option in Google's lineup rather than against flagship models.

Where is Google already using it?

Google says Gemini Notebook runs the API as a real-time router for model tiers and thinking levels, delivering up to 40% lower latency for everyday queries while reserving deeper reasoning models for complex tasks. That deployment is the clearest statement of intent: Google treats decision routing as a distinct workload with its own economics, not a job for the general-purpose chat model.

Two other internal deployments appear in the announcement. Stitch's rapid canvas experiment uses the API to generate playable wireframes. CC from Google Labs uses it to select among more than 200 emoji reactions — a ranking task where the model must pick one option from a large fixed set, exactly the shape of problem the choice question type addresses.

How do developers get access?

Developers can call the endpoint from Python, JavaScript or REST. The SDK quickstarts require google-genai 3.2 or later. A no-cost AI Studio tier supports prototyping.

Third-party integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode. Gemini in Google Cloud support is listed as coming soon.

The current release carries limits. Initial access covers only the Gemini Developer API v1beta with Gemini Decision 1. Question instructions are text-only. Conversation history and tool-call blocks are unsupported. Context caching, the Batch API, and Flex or Priority inference are unavailable.

Who owns the risk when the model decides?

The API consolidates multimodal judgments into a single typed call, but it does not take on the safety burden that follows. Google places responsibility for safe automation with the developer — specifically with the criteria written into each question, the labeled evaluation data used to calibrate confidence thresholds, and the escalation rules that decide when a machine answer must not act alone. The documentation's own example of payouts as a high-stakes action signals where Google draws the line: the model supplies the judgment, the developer supplies the guardrails.

The launch lands as agent frameworks multiply the number of small routing decisions per user interaction, and per-decision cost and latency compound. A dedicated scoring endpoint at $0.10 per million input tokens, with flat latency across question count, gives agent builders a way to move that control flow off the chat path entirely. What it does not change is the audit question: every automated branch still traces back to criteria a human wrote.