Wain AI/Tech Blog

AI news and trends worldwide, updated nearly every day

Google Launches the Gemini Decisions API, Which Returns Judgments Instead of Text - No Output Charges, $0.10 per Million Input Tokens

Google Launches the Gemini Decisions API, Which Returns Judgments Instead of Text - No Output Charges, $0.10 per Million Input Tokens

On November 11, 2026, Google released the Gemini Decisions API and a new model, Gemini Decision 1, which answer yes/no, choice, and scoring questions with probabilities and confidence values. Output tokens are always zero, and input costs $0.10 per million tokens. This article covers how billing works, where it fits, and what it does not support.

On November 11, 2026, Google added a new endpoint to the Gemini API, the Decisions API, along with a model built for it, Gemini Decision 1 (gemini-decision-1)1. What comes back is a probability rather than prose. The API only takes questions whose possible answers are fixed in advance: is this message reporting a damaged item, which queue should it go to, which of three severity levels applies.

Input may mix text, images, audio, video, and PDF files, and one request can carry up to 1,000 questions. Output tokens are always zero, so on the paid tier the only charge is $0.10 per million input tokens3. Teams that currently build agent routing or tool selection out of a generative model call followed by JSON parsing now have one more candidate to swap in.

Probabilities for Three Kinds of Questions, With No Text Written

The documentation frames the problem this way: generative models produce text meant for people, and producing text just so a program can parse a branching decision out of it costs time and breaks easily2. Instead of producing an answer one token at a time, Gemini Decision 1 scores the candidates in one forward pass1.

There are three question types.

TypeWhat you can askWhat comes back
noulA yes/no judgmentProbability of yes (0.01 to 0.99) and confidence
choiceOne pick from 2 to 255 optionsThe top option, probabilities for every option, confidence
scoreA rating on an ordered rubric of 2 to 255 levelsProbability-weighted average of the levels, per-level probabilities, confidence

Every question is evaluated on its own against the shared input and cannot see the others. Adding or dropping a question leaves the remaining answers unchanged2. For judgments that depend on each other, such as choosing a repair category only when an item is damaged, the guidance is to make a follow-up call once the first result is in, or to ask both together and discard whichever result turns out to be irrelevant.

Confidence, on a scale from 0 to 1, estimates how likely the leading answer is to be right. In the documentation’s example, answers at 0.7 or above are routed automatically and anything lower goes to human review. Google asks developers to set that threshold from labeled examples of their own traffic, weighing what a false positive costs against a false negative2.

The numbers carry a few traps. A yes/no probability always stays between 0.01 and 0.99, so a condition written as “above 0.99” will never fire. A choice question spreads all of its probability across the listed options even when none of them fits, which means you need a confidence threshold, a separate yes/no check on scope (for example, whether a message concerns an order at all), or an extra option like “other” to absorb the misfits. Before evaluating, the API reorders option keys alphabetically2.

Billing Covers Input Only, but Caching and Batch Are Off the Table

Costs add up differently from a generative model. However many questions ride along, the shared input is charged a single time per call; each question contributes only the tokens of its own wording, option descriptions, and formatting3. In its announcement, Google says you pay nothing for output tokens, face no unpredictable bills for thinking tokens, and no extra fees for audio transcription or OCR1.

ItemFree tierPaid tier (per million tokens)
Input (text, image, video, audio, PDF)Free$0.10
OutputNot applicable (0 tokens)Not applicable (0 tokens)
Context cachingNot offeredNot offered
Used to improve Google’s productsYesNo

On the other hand, context caching, the Batch API, and Flex and Priority inference are all unsupported. Input that repeats across requests is counted in full every time2. If you place a long policy document in every request and only change the questions, packing more questions into one request lowers the cost per question, and splitting them across requests raises it. Because free-tier input is used to improve Google’s products, running business data through the API realistically means using the paid tier3.

On speed, Google says probabilities and confidence values come back in 150 to 250 milliseconds end to end1. That is the company’s own figure, and only measurement will show how your inputs behave.

Where It Fits and Where It Doesn’t

The uses Google lists include routing requests to models or tools, screening content against a large set of policy rules in one call, deciding an agent’s next step (such as which tool to use or whether to stop), and scoring candidates2. Inside Google, Gemini Notebook reportedly uses it as a router that sends each query to a suitable model and thinking level1.

Generating prose, summaries, or explanations, pulling values that match your own JSON schema, calling tools with arguments, and multi-step reasoning stay with the Interactions API, according to the documentation’s comparison table2. Input is restricted as well: conversation history, thought blocks, and tool-call blocks are rejected, so multi-turn context has to be flattened into text.

Two of the documentation’s cautions look especially relevant in practice. The first is to treat input as untrusted. Text that users write may try to talk the model into a favorable label, so Google recommends explicit criteria, testing with adversarial examples, and keeping a person or a second check in front of high-impact actions such as payouts2. Because the result feeds straight into a branch in your code, defenses against prompt injection are likely to be just as necessary here as with a generative model. The second is to handle arithmetic on figures, dates, and time spans in your own code and hand over just the result or a labeled bucket. The documentation also recommends trimming the input down to what the questions require.

Availability is still narrow. For now it is offered solely through the v1beta Gemini Developer API, with SDK 3.2 or newer required2. Support for Gemini in Google Cloud is described as coming soon, and integrations with OpenRouter, LangChain, Vercel, LiteLLM, and others are said to be rolling out1. Organizations that run on Google Cloud would, for now, have to test it through the Gemini Developer API instead.

Choosing Between It and October’s Strands Decider 2B

A model that returns judgments without generating text also appeared as open source in October. Strands Decider 2B, released by strands-labs on October 1, is a small open source model that runs on a local CPU or GPU.

The choice between them comes down to where it runs and how much it handles. Strands Decider 2B has published weights and runs in your own environment. The Decisions API runs on Google’s cloud and accepts up to 262,144 tokens of input, with image, audio, video, and PDF content allowed2. Both return a confidence value for every judgment and show how to hand low-confidence cases to a person.

A quick way to evaluate it is to pick one place where an LLM call currently drives a branch and the possible answers are fixed. Label the judgments your existing pipeline produced, then count agreement and the number of cases sent to people at each confidence threshold. That shows, on your own data, how latency, cost, and accuracy would change with the swap.

Sources

  1. Gemini Decisions API: Fast, calibrated multimodal decisions - Google official blog (November 11, 2026)
  2. Decisions API - Gemini API documentation (question types, limits, when to use it; checked November 12, 2026)
  3. Gemini Developer API pricing - Gemini API documentation (Gemini Decision pricing; checked November 12, 2026)

We publish the latest AI news nearly every day.

Subscribe via RSS Get new posts the moment they go live.

Search other keywords →