Takeaway: On November 11, 2026, Google DeepMind introduced the Gemini Decisions API, a new Gemini API endpoint powered by a new model, Gemini Decision 1 (gemini-decision-1), in Google AI Studio. Instead of generating an answer token by token, the model scores a shared multimodal input against up to 1,000 independent yes/no, multiple-choice, or rubric questions in one forward pass and returns calibrated probabilities and confidence scores, with zero output tokens billed. Google’s pricing page lists a free tier with input free of charge, and a paid tier of $0.10 per 1 million input tokens for text, image, video, audio, and PDF. Batch API and context caching are not available. This is a developer control-plane launch, not a consumer Gemini app change, and it is separate from Gemini 4 Carbon’s November 10 access and price post and from Gemini arriving on Apple CarPlay.
The official Decisions API documentation is the right place to start, alongside the Gemini Decision 1 model page. The docs require the Google GenAI SDK 3.2 or later, and calls go through the v1beta API. A typical request names the model, passes shared input, and sends a map of questions. Answers come back under the same keys, with usage that reports input tokens by modality and total_output_tokens of 0.
Why Google built a decision endpoint instead of another chat model
Google’s framing is straightforward. Generative models are built to write text people can read. A lot of production code does not need a paragraph. It needs a branch: which tool to call, which queue should get the ticket, whether a policy rule fired, or how a candidate scores against a rubric. Parsing that judgment out of generated prose adds latency and brittle glue. The Decisions API is meant to skip that loop. Decision 1 scores the options you already defined and returns numbers your code can threshold.
The documentation draws a sharp line against the Interactions API. Use Decisions when you want a yes/no probability, one option from a fixed set, a rating on a fixed scale, many independent checks over the same input, low latency, and no output-token bill. Use Interactions when you need generated text, summaries, explanations, structured JSON of your own design, tool calls with arguments, or multi-step reasoning. If you find yourself asking Decision 1 to “explain why,” you are in the wrong endpoint.
Three question types in one request
Every question is one of three types. You can mix them in a single call. Each question is scored independently against the shared input; questions cannot see each other, and the model never sees your question keys, so the full question has to be written in the instructions field.
- Yes/no (
noul). Returns the calibrated probability that the answer is yes, plus a confidence score. Optional criteria can define what yes and no mean. The docs say yes/no probabilities always fall between 0.01 and 0.99, so a threshold above 0.99 will never match. - Choice. Selects from 2 to 255 options, returning the winner, a probability for every option (they sum to 1), and confidence. Criteria map each option ID to a description; use an empty string when the ID is self-explanatory. Option keys are sorted alphabetically in UTF-8 byte order before evaluation, so “none_of_the_above” is a bad catch-all name — the docs recommend a non-positional key such as “other.”
- Score (ordered rubric). Grades the input across 2 to 255 ordered levels, indexed from 0. The docs recommend 2 to 10 levels for most rubrics. The returned score is a probability-weighted average of those indices, so it can land between levels. The response also includes per-level probabilities, a legend, and confidence. Because different distributions can share the same expected value, the docs tell you to look at the probability map, not only the average, and to take the highest-probability level when you need a discrete bucket.
Confidence is the estimated probability, from 0 to 1, that the top answer is right. For yes/no questions it equals max(noul, 1 - noul). Google’s own examples route low-confidence results to a human or a fallback. That is the intended control pattern: the value picks an action, and confidence decides whether to run it automatically.
Shared multimodal input, billed once
Decision 1 accepts text, images, audio, video, and PDFs in the same Content block format used by the Interactions API, up to a 262,144-token context window across the shared input and all questions. You can combine modalities in one input — the announcement’s example is a call recording plus account metadata, or a scanned PDF claim plus damage photos. Image and video blocks can set a media resolution (low, medium, high, or ultra_high; ultra-high is images only) to trade visual detail against token cost.
The billing rule is the one that makes a 1,000-question request interesting: shared input is counted once per request, no matter how many questions you attach. Each question adds only its own instruction and criteria tokens. There is no output-token charge, no separate audio-transcription or OCR line item, and no thinking-token bill, because the model is not decoding an answer. Video frames and PDF page images are counted as image tokens in the usage breakdown.
The announcement puts median model compute at 50–100 milliseconds and end-to-end latency at 150–250 milliseconds for multimodal inputs and contexts up to 16K tokens, “remaining nearly flat from 1 to 100 questions per request.” That latency claim is Google’s, and it is scoped to those context sizes. The context window itself goes much higher, to 262,144 tokens. Do not treat the 150–250 ms band as a guarantee at the top of the window.
Price, free tier, and hard limits
The official price list is simpler than most Gemini model cards. On the paid tier, input is $0.10 per 1 million tokens for text, image, video, audio, and PDF. Output price is listed as not applicable, because output tokens are always 0. Context caching is not available. The free tier lists input as free of charge, also with no output-token price. As with other Gemini free-tier rows, the pricing page says free-tier usage may be used to improve Google’s products; paid-tier usage is listed as not used that way. Confirm the live price page when you budget. The November 11 blog post also mentions “an API no-cost tier you can use in Google AI Studio for prototyping,” which matches that free-tier row.
The docs are equally explicit about what the endpoint will not do:
- It is available on the Gemini Developer API (v1beta) with
gemini-decision-1only. The announcement says support is coming soon to Gemini in Google Cloud; that is a future path, not a current one. - Question instructions are text only. You cannot put media inside a question.
- YouTube URLs and
gs://Cloud Storage URIs are not supported. Media has to come through the Files API, a public or signed HTTPS URL, or inline base64. - Conversation history, thought, and tool-call blocks are not valid input. Multi-turn context has to be serialized as text.
- Context caching, the Batch API, and Flex or Priority inference are not supported. Repeated input is billed in full each time.
- A request can include 1 to 1,000 questions, 2 to 255 choice options, and 2 to 255 score levels, all still bounded by the 262,144-token window. Question keys are 1 to 64 ASCII characters matching
[A-Za-z0-9_-]+. - If input safety filters block a request, the API returns HTTP 400 with a JSON error, not a partial
answersobject.
Rate limits are documented on a separate Gemini Decision rate-limits section. We are not restating quota numbers here, because those tables change; check the live page for your tier.
How Google says it is already using the API
The company post lists four internal examples. Treat them as Google’s own illustrations, not as independent case studies:
- Gemini Notebook uses the API as a real-time router, sending everyday queries to a cheaper or faster tier and reserving deeper reasoning models for harder work. Google says this can serve everyday queries with up to 40% lower latency. That is a company figure.
- A rapid-canvas experiment in Stitch uses the API to generate playable wireframes and apply voice or tap edits quickly, then hand the screens to the full Stitch experience.
- 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 within the latency budget.
The announcement also points developers to Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper (the docs also call the last one Find the Zookeeper) demo apps in Google AI Studio, plus a Decisions Playground for interactive testing. Ecosystem integrations are “rolling out” across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode, alongside the google-genai SDKs. “Rolling out” is not the same as “already supported in every one of those tools today,” so verify the integration you plan to use.
Benchmark claims — and how to read them
Google says Decision 1 leads a suite of benchmarks covering agent routing, reasoning, ranking, and multimodal understanding, with the lowest calibration error in that suite, and that Decision 1 with the Decisions API is nearly 5x cheaper across the shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API. Those are Google-selected comparisons, run on tests Google is citing. We have not rerun the harnesses. A 5x cost claim that depends on zero output tokens and on a particular mix of tasks will not automatically hold if you still need a generative model to write the user-visible answer after the decision, or if your questions are so numerous that instruction tokens dominate the bill.
The right way to evaluate the endpoint is narrower: pick the closed-set decisions you already make in code, send them with explicit criteria, threshold on confidence, and compare quality, latency, and cost against your current classifier or “ask a flash model and parse JSON” loop. The docs themselves warn you to treat user-supplied input as untrusted, because text can argue for its own classification, and to keep a human or a second check on high-stakes actions such as payouts.
How this sits next to other Gemini news this week
Carbon’s November 10 post opened a long-horizon generative model to the paid API, Antigravity, and Ultra surfaces, at a very different price ($2 / $10 through November 30, then $4 / $20 for the public-access model). Decision 1 is a cheaper, faster, non-generative scorer. You might use Decision 1 to decide whether a request should go to Carbon, to Argon, or to a smaller model — that is literally one of the routing examples Google gives — but you would not replace Carbon with Decision 1. Gemini on Apple CarPlay is a consumer Live surface. Maps Platform in AI Studio is a prototyping path. None of those launches is part of the Decisions API.
API facts at a glance
| Item | What Google’s November 11 announcement and docs say |
|---|---|
| Endpoint and model | Gemini Decisions API, model ID gemini-decision-1, Gemini Developer API v1beta |
| What it returns | Calibrated probabilities and confidence for yes/no (noul), choice, and rubric-score questions; zero output tokens |
| Input | Text, images, audio, video, and PDFs; shared across all questions in the request |
| Scale | 1 to 1,000 questions per request; 2 to 255 choice options or score levels; 262,144-token context window |
| Latency (Google’s claim) | 150–250 ms end-to-end, 50–100 ms median model compute, for contexts up to 16K tokens and up to 100 questions |
| Paid price | $0.10 per 1 million input tokens; output not applicable; shared input billed once |
| Free tier | Input listed as free of charge for prototyping; see the live price page and terms for data-use differences |
| Not supported | Batch API, context caching, Flex or Priority inference, YouTube or gs:// URIs, media inside questions, Google Cloud (coming soon, per the announcement) |
| SDK | Google GenAI SDK 3.2 or later; REST endpoint at https://generativelanguage.googleapis.com/v1beta/decisions |
FAQ
What is the Gemini Decisions API?
It is a Gemini API endpoint, launched on November 11, 2026, for closed-set decisions. You send shared input and a map of yes/no, choice, or rubric questions. Gemini Decision 1 scores the options in one pass and returns probabilities and confidence scores instead of generated text.
How much does Gemini Decision 1 cost?
On the paid tier, Google lists $0.10 per 1 million input tokens for text, image, video, audio, and PDF, with no output-token charge. Shared input is billed once per request; each question adds its own instruction and criteria tokens. A free tier lists input as free of charge. Context caching and Batch are not available. Confirm the live pricing page before you forecast a production bill.
Is this a replacement for Gemini 4 Carbon or for the regular Gemini API?
No. Decision 1 does not generate text, call tools, or reason through a long job. Carbon and the Interactions API still cover those jobs. The Decisions API is for bounded judgments you can express as yes/no checks, a fixed list of options, or a rubric.
Can I ask 1,000 questions about a PDF or a video in one call?
Yes, up to 1,000 questions, as long as the shared input plus every question’s instructions and criteria fit in the 262,144-token window. Questions are scored independently. If one decision depends on another, either send a second request or ask both and ignore the answer you do not need.
What does “noul” mean?
It is Google’s name for the yes/no question type in this API. The field returned on that answer is also called noul, and it is the calibrated probability that the answer is yes, bounded between 0.01 and 0.99.
Does the API work with images, audio, and video?
Yes. Decision 1 is multimodal. You can mix text, images, audio, video, and PDFs in the shared input and set image or video resolution to control token cost. Media cannot be placed inside a question’s instructions, and YouTube or Cloud Storage URIs are not accepted.
Is Gemini Decision 1 available on Google Cloud or in Vertex AI today?
The docs say the endpoint is available on the Gemini Developer API (v1beta) with this model only. The November 11 announcement says support is coming soon to Gemini in Google Cloud. That is a planned expansion, not a current availability claim.
Should I trust the 5x cheaper and 150–250 ms figures?
Those are Google’s figures, from the company’s announcement: a nearly 5x lower cost across a shared benchmark suite versus Gemini 3.5 Flash-Lite with the Interactions API, and 150–250 ms end-to-end latency for contexts up to 16K tokens. They are useful as the company’s stated positioning. They are not a substitute for a test on your own questions, traffic mix, and confidence thresholds.