Editorial

Deep dive The Forward Pass research desk

Gemini Decisions Separates Scoring From Generation

Google’s new endpoint scores bounded choices without generating answer text, with shared-input billing and latency claims that engineers should distinguish from its maximum request limits.

AI-generated article

This article is researched and written by AI models from the sources it cites, and every claim is checked against them. No human edits it before it is published.

Kind
Deep dive
By
The Forward Pass research desk
Published
November 15, 2026
Read time
5 min

What shipped

Google has introduced the Gemini Decisions API, an endpoint for closed-set decisions powered by Gemini Decision 1, identified as gemini-decision-1. It is available through the Gemini API in Google AI Studio, including a no-cost API tier for prototyping.

The distinction is the output contract. Google says the model scores candidate options in a single forward pass rather than generating an answer token by token. The endpoint is positioned for bounded choices such as tool selection, policy checks, model routing, and context grading.

If your application needs a choice among defined alternatives, this is the interface Google is offering. If it needs generated answer text, that is not the operation described in this launch. Evaluate it as a scoring endpoint rather than assuming it is a drop-in replacement for a text-generation call.

Google supports shared inputs combining text, images, audio, video, and PDFs. A request can contain up to 1,000 independent questions, with an advertised input limit of up to 256K tokens.

Those are capacity limits, not the boundaries of every performance claim. Google’s nearly flat latency result covers a smaller range of question counts and context lengths.

How the decision interface works

Google describes three supported question types:

  • Yes/no probability estimates.
  • Multiple-choice questions.
  • Ordered rubric scores.

According to Google, every answer includes a confidence score from 0 to 1. The company explicitly suggests using that score to route ambiguous cases to a fallback or human reviewer.

A policy check might fit a yes/no question, while tool selection is among the bounded-choice use cases Google names. Google also lists context grading and model routing among its intended uses.

The request model also deserves attention: Google describes the questions as independent and evaluated against a shared multimodal input. Treat that as the documented unit of work. Do not read the maximum question count as a description of a sequential workflow in which later questions consume earlier answers.

Confidence is an additional output to evaluate, not a supplied operating threshold. Google reports strong calibration results, but the provided comparison does not specify a confidence cutoff for your routing, policy, or review requirements.

A useful evaluation should therefore examine both the selected answer and the confidence attached to it.

The latency claim has a defined envelope

Google reports median model-compute latency of 50–100 ms and end-to-end latency of 150–250 ms. It attributes the execution model to scoring options without token-by-token decoding.

The more specific scaling claim is that latency remains nearly flat from 1 to 100 questions per request, across multimodal inputs and contexts up to 16K tokens. That is the reported performance envelope to use when interpreting the headline numbers.

It is not the same as the endpoint’s advertised maximum of 1,000 questions and 256K input tokens. The supplied figures do not establish nearly flat latency across those larger limits. Nor should the 50–100 ms model-compute figure be substituted for the separately reported end-to-end figure.

For an evaluation, keep the variables explicit:

  • Question count, including whether it falls within the reported 1–100 range.
  • Shared-input length, including whether it stays within 16K tokens.
  • The modalities present in the request.
  • Whether the measurement is model compute or end to end.

Google’s result is relevant if your workload sits inside that envelope. If you need the larger request limits, measure that workload directly rather than extending the reported curve by assumption.

The launch supplies an aggregate latency claim, not a separate guarantee for every combination of modality, question type, context length, and question count.

What the price actually covers

Google lists Gemini Decision 1 at $0.10 per million input tokens and $0 for output tokens. The important billing detail is how it treats a shared input: that input is billed once per request regardless of how many questions the request evaluates.

Each question still adds its own instruction and criteria tokens. More questions therefore do not mean repeatedly billing the shared input, but neither does the pricing description make additional questions token-free.

That distinction should shape a cost test. Account for the shared material and the question-specific material separately.

Zero output-token charges also need a narrow reading. They are part of this endpoint’s pricing, not a claim that the overall decision costs nothing. Input tokens remain billable, including the instructions and criteria attached to individual questions.

Google offers a no-cost API tier for prototyping in AI Studio. The supplied facts do not specify its allowance or other tier limits, so the existence of that tier is not enough to construct a production cost estimate.

The published input rate and shared-input rule provide the pricing basis. The benchmark comparison is a separate claim about a particular evaluated workload.

What the benchmark comparison establishes

Google says Gemini Decision 1 with the Decisions API is nearly five times cheaper across its shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API. The comparison names both a model and an API on each side.

Keep that pairing intact. This is not a stated comparison against every Gemini model, every API interface, or every possible classification workload. It is Google’s reported cost result across the shared suite it evaluated.

Google also says Gemini Decision 1 leads on a benchmark suite spanning agent routing, reasoning, ranking, and multimodal understanding, with the lowest calibration error. These are Google’s claims, not independent evaluation results.

The supplied facts do not include per-task scores, calibration-error values, or a workload breakdown sufficient to reproduce the headline cost comparison. There is consequently no supported numerical accuracy margin to carry into an application decision, and no basis here for applying the nearly fivefold cost result universally.

That is especially important if confidence determines escalation. Google reports the lowest calibration error across its benchmark suite, but the supplied facts do not include an escalation threshold.

Who should evaluate it now

The clearest candidates are teams building the bounded-choice workflows Google names: tool selection, policy checks, model routing, and context grading. The API’s supported question types give those teams a concrete interface to test without asking for generated answer text.

Availability also needs a precise reading. Google says integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode, alongside the google-genai SDKs. That wording does not establish that every named integration is already available for your chosen path.

Support for Gemini in Google Cloud is described as coming soon. Teams that require that access route should distinguish it from the currently announced availability through the Gemini API in Google AI Studio.

For a direct start, Google points developers to an AI Studio API key and Python, JavaScript, or REST guides. The linked demo apps are Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper.

Google’s launch makes the interface and pricing explicit. Its broader benchmark claims remain claims to test against the workload you intend to ship.

Try it

Start in Google AI Studio by getting an API key, then follow Google’s Python, JavaScript, or REST Decisions API guide with model gemini-decision-1. No executable commands were supplied in the source material, so there are no command steps to reproduce here.

Sources