Gemini Decision 1: Multimodal Decisions and API Pricing

A yes/no question uses the API type noul and returns the probability that the condition holds. A choice question selects from 2 to 255 named options and returns their probabilities. A score question uses ordered descriptions starting at level zero and returns a probability-weighted level, so the result can fall between two rubric levels.

Define options in terms of the decision your application can actually take. For ticket routing, distinguish billing, technical support and an explicit other route. A generic instruction such as “is this good?” leaves the model to infer your standard; a rubric describing missing evidence, a usable answer and a complete answer makes the decision testable.

Each question sees the shared input, but it cannot see another question’s answer. If a later action depends on a previous result, combine independent answers in code or make a second request. Use a generative model for explanations, summaries or rewritten content; this endpoint returns typed decisions.

Create a Gemini API key in Google AI Studio and make sure the project can access gemini-decision-1. The REST endpoint uses the Gemini Developer API v1beta. Google’s SDK quickstart requires version 3.2 or later of google-genai; the request below uses REST so the question format is visible.

Start with a labeled input and inspect answers, confidence and usage before connecting the result to an automated action. This example asks two independent questions about the same ticket.

The paid rate is $0.10 per million input tokens across supported modalities, with $0 output tokens. Count the shared content plus every question’s instructions and criteria. The shared input is billed once within a request, regardless of the question count; adding questions still adds their own tokens. The free tier has separate usage limits and allows data to be used to improve Google products; paid-tier inputs are not used for that purpose. Choose the appropriate tier before sending customer material.

For example, a request containing 10,000 shared-input tokens and 1,000 total question tokens costs $0.0011. At 100,000 identical-size requests, the token cost is $110. If ten separate calls each repeat that shared input and use 100 question tokens, they total 101,000 tokens and cost $0.0101 instead of $0.0011 for the combined request.

These are token-budget examples, not media-to-token conversions. Read the returned usage for actual images, recordings and documents. Repeated inputs across separate requests are billed in full: this API does not support context caching, Batch, Flex or Priority inference.

The 262,144-token context window covers both shared input and all questions. The 1,000-question maximum is also bounded by that total: long documents and large option lists leave less room for additional questions. Question instructions are text only; put image, audio, video and PDF blocks in the shared input.

Media can use a Files API URI, a public or signed HTTPS URL, or inline base64. Direct gs:// and YouTube URLs are not supported. The documentation lists a 15MB limit per external URL, 100MB for inline request data and 50MB for inline PDFs. Set each block’s MIME type to the actual file format.

A call recording plus account notes can support separate urgency, intent and next-team questions. Keep unrelated customer records in separate requests: shared-input scoring does not give each question its own isolated document. Serialize conversation history as text because thought and tool-call blocks are not accepted.

Choice probabilities are distributed across the options you supply even when none fits. Include an explicit fallback, or pair routing with a separate scope check. For yes/no answers, probabilities range from 0.01 to 0.99; a threshold above 0.99 will never match. Choose action thresholds using labeled cases and the cost of an incorrect action.

Google reports 150–250ms end-to-end latency for its launch evaluation, with model-compute measurements covering inputs up to 16K tokens and near-flat behavior from 1 to 100 questions. Those conditions do not establish the latency of a full 262K-token, 1,000-question request. Measure the complete path from your application, including media transfer and retries.

In a November 12 Reddit report, the author tested 54 free-tier calls and found that an off-topic ticket was confidently routed to technical support when the choices lacked an “other” fallback option. Include out-of-scope inputs in your test set alongside labeled examples.

Choose Gemini when the shared input includes audio, video or PDFs alongside text and images, or when you need up to 1,000 independent questions within one context window. Decider v1.3 and Flash accept text and images and add locate questions for verbatim text extraction. A Decider request containing locate is text-only, allows at most 128 questions and cannot use web grounding.

At the published direct API rates, 20 million billed input tokens cost $2 with Gemini, $0.40 with Decider v1.3 or $0.20 with Flash v1.3. All three have free output. These are equal-billed-token examples: Gemini bills shared input once per request, while Decider can split questions into groups and bill the state again for each group. Use returned usage totals when comparing a large question set.

Decider also publishes downloadable checkpoints; Gemini Decision 1 is hosted. Test both on the same labeled cases and compare incorrect routes, fallback rates and full request latency. A lower input rate only helps if the chosen model meets the required decision quality.

Gemini Decision 1 is a hosted option when one input includes several media types or many independent questions. Microsoft-Decision-1 offers a hosted text-only alternative. Decision 3.0 offers local multimodal weights when control of deployment matters. Compare label accuracy and fallback rates on the same cases before comparing the token bill or instance hours.

Gemini 4 Carbon serves a different part of the workflow: generating code, analysis and explanations. A decision endpoint can select a route before a generative call, but it does not execute the selected tool or manage an agent’s permissions.