88Z AI, 12 November — On 11 November 2026, Google DeepMind introduced the Gemini Decisions API, powered by a new model, Gemini Decision 1 (gemini-decision-1). It is built for bounded judgments such as yes/no, multiple choice, and rubric scoring. One request can ask up to 1,000 independent questions about a shared multimodal input and return probabilities and confidence scores. The announced price is $0.10 per million input tokens, with output tokens billed at zero.
At a glance
- The API scores closed-set options in one forward pass instead of generating answer text token by token.
- A request can include up to 1,000 independent questions, mixing yes/no, multiple choice, and rubric scores, each with a confidence value from 0 to 1.
- Google cites 150–250 ms end-to-end latency and a price of $0.10 per million input tokens, with output billed at zero.
- It fits routing, policy checks, and grading; ambiguous results should have a confidence threshold and a fallback or human review.

The job is a batch of bounded choices, not another stretch of prose
On 11 November 2026, Google DeepMind product manager Min Choi and principal scientist Felix Yu introduced the Gemini Decisions API. It is available through the Gemini API in Google AI Studio, and is powered by a new model, Gemini Decision 1 (gemini-decision-1). Officially, it provides primitives for bounded choices such as tool selection, policy checks, model routing, and context grading.
Unlike ordinary chat completion, it does not generate an answer token by token. It scores the candidate options you supply in a single forward pass and returns probabilities and confidence scores. You can send one multimodal input of up to 256K tokens, including text, images, audio, video, or PDFs, plus up to 1,000 independent questions. Official examples include pairing a call recording with account metadata, or a scanned claims PDF with photos of damage.
Three question types, plus the official latency and price
Questions come in three types. noul is a yes/no check, with optional descriptions of the true and false cases, and an estimate of the probability that the condition holds. choice selects among 2 to 255 options and returns the winner plus normalized probabilities for every candidate. score grades an input on an ordered rubric starting at 0, returning a continuous weighted score and per-level probabilities. Every answer includes a confidence score from 0 to 1, so your code can send ambiguous cases to a fallback or a human reviewer. The input field uses the same Content block format as the Interactions API.
On latency and cost, the post cites 150–250 ms end-to-end, with 50–100 ms of median model compute, remaining nearly flat from 1 to 100 questions per request at contexts up to 16K tokens. Gemini Decision 1 is priced at $0.10 per million input tokens and $0 for output tokens. The shared input is billed once per request, and each question adds only its own instruction and criteria tokens. Google also says there are no output-token charges, variable thinking-token bills, or separate audio transcription or OCR costs. Those are announcement prices, not 88z.ai or another platform’s retail rates. The post also mentions a no-cost API tier in Google AI Studio for prototyping.
How to start, and which official examples were given
Developers can get an API key in Google AI Studio and follow the Python, JavaScript, and REST guides in the Decisions API documentation. The sample in the post uses the google-genai client, sets the model to gemini-decision-1, passes an image and text together as input, and mixes question types in the questions object. Official demo apps include Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper.
On the product side, the post mentions several internal uses: CC from Google Labs uses it to pick an emoji reaction from a palette of more than 200 options; Google Search uses it as a safety layer for MCP tools to detect adversarial behavior; and Stitch Rapid uses it to design a product flow in seconds. Ecosystem integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode, alongside the google-genai SDKs, with support coming soon to Gemini in Google Cloud. Items described only as coming soon should not be treated as already live on every platform.
88Z perspective: use it on genuinely closed questions, and leave a path for low confidence
If your system already uses a large model for routing, policy checks, or grading, and you often struggle to parse a model that prefers to write an explanation, this API’s design is worth a close look: shrink the question to yes/no, a finite set of choices, or a rubric, process a batch at once, and read the confidence score. It does not replace work that needs open-ended reasoning or long-form writing. The official claims of benchmark leads and lower calibration error come from Google’s own evaluation, which this site has not repeated. You should still test accuracy, calibration, and cost on your own samples.
A practical start is one clearly bounded task, such as whether a ticket belongs to billing, with a confidence threshold below which the case goes to a person or a stronger chat model. Also check whether your inputs truly belong in one shared context, and whether the 256K-token and 1,000-question limits cover your batch. Prices and the free tier should be taken from Google’s current docs. If you call it through 88z.ai or another platform, confirm separately that the endpoint is actually offered; do not assume every OpenAI-compatible platform already exposes the Decisions API.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- Google DeepMindSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.