Gemini Decisions API: gemini-decision-1 scores yes/no and choice questions at $0.10 per million tokens

Google has released the Gemini Decisions API, a developer endpoint that does not write text at all. Instead of generating an answer, the model gemini-decision-1 reads an input and scores a list of questions about it, returning probabilities and a confidence value for each. Google announced it on 11 November 2026 and prices it at $0.10 per million input tokens with no charge for output. For developers who use a large language model mainly to branch, route or filter, it is a different and cheaper tool for that job.

Key points

• Google's Gemini Decisions API, announced 11 November 2026, uses the model ID gemini-decision-1 and returns scores instead of generated text.
• Pricing is $0.10 per 1M input tokens, $0 for output, with a free tier in Google AI Studio.
• One request can carry up to 1,000 independent questions of three types: yes/no (noul), choice and score.
• Google says end-to-end latency is 150 to 250 ms and the API is nearly 5x cheaper than Gemini 3.5 Flash-Lite on its shared benchmark suite.
• It is in v1beta on the Gemini Developer API only; Gemini in Google Cloud support is listed as "coming soon".

What happened

On 11 November 2026 Google published a post by Min Choi (product manager) and Felix Yu (principal scientist) of Google DeepMind introducing the Gemini Decisions API. Google describes it as a closed-set decision endpoint: you give it one input and a set of questions, and a single forward pass of the model scores every question. Nothing is generated token by token, so the output token count is always zero, as the API documentation states.

The input can be text, images, audio, video or PDFs, in the same content-block format as Google's Interactions API. The docs give a combined context window of 262,144 tokens across the input and all questions. A request holds between 1 and 1,000 questions, and each question is scored independently against the shared input. The model never sees the question keys, which are only labels for you.

There are three question types:

• noul (yes/no): returns the probability that a condition holds. The docs say this value stays between 0.01 and 0.99, so a threshold above 0.99 will never match.
• choice: you supply 2 to 255 options with descriptions. The answer is the top option plus normalised probabilities for all options. Option keys are sorted alphabetically in UTF-8 byte order before evaluation and ties go to the alphabetically first key.
• score: you supply an ordered rubric of 2 to 255 levels (the docs recommend 2 to 10). The answer is a probability-weighted score, the per-level probabilities and a legend.

Every answer also carries a confidence from 0 to 1, which the docs define as the estimated probability that the top answer is correct. The intended pattern is to act automatically above a threshold and send the rest to a fallback model or a person.

The Python call in the docs is short: create a genai.Client with api_version set to v1beta, then call client.decisions.create(model="gemini-decision-1", input=..., questions={...}) and read result.answers["is_damaged"].noul.

Background

Most production uses of an LLM that are not chat are decisions in disguise: is this message spam, which tool should the agent call, is this retrieved passage relevant, does this request break policy. Developers usually solve them by prompting a general model to answer in JSON, then parsing the text. That works, but it spends output tokens, adds latency, and gives no calibrated sense of how sure the model was.

Google's own recent guidance points the same way. We covered its advice to route work with cheap classifiers across Gemini models and thinking levels, and our piece on comparing LLM prices by cost per task explains why output-heavy routing steps are a hidden cost. The Decisions API is a dedicated version of that router.

The numbers

Google's pricing page lists gemini-decision-1 at $0.10 per 1M input tokens on the Standard tier for text, image, video, audio and PDF input. Output is not applicable. The Free tier is free of charge, but the page says free-tier content is used to improve Google's products and paid-tier content is not. Context caching and the Batch API are not supported.

Billing has one feature worth noting. Shared input is billed once per request regardless of the number of questions, and each question adds only the tokens of its own instructions and criteria. Ask 50 questions about one long document and you pay for the document once. Google also says there are no charges for thinking tokens or for separate audio transcription or OCR.

On speed, Google says end-to-end latency is 150 to 250 ms, with median model compute of 50 to 100 ms, measured for contexts up to 16K tokens, and that latency stays nearly flat from 1 to 100 questions per request. On quality, the company says Gemini Decision 1 leads its benchmark suite (agent routing, reasoning, ranking and multimodal understanding) and has the lowest calibration error, and that the API is nearly 5x cheaper than Gemini 3.5 Flash-Lite used through the Interactions API on that suite. The post gives no individual benchmark scores in the text we read, and these are Google's own comparisons that we could not reproduce.

Google also names uses inside its own products: Gemini Notebook routes queries to the right model tier and thinking level, with "up to 40% lower latency for everyday queries", according to the post; a Google Labs product called CC picks an emoji reaction from a palette of more than 200; and Google Search uses it as a safety layer to detect adversarial behaviour in MCP tools.

How it works

Think of it as a classifier you configure with plain-language descriptions instead of training data. A routing example: input is the user's request, and one choice question lists the options "answer_directly", "search_web" and "escalate_to_human", each with a one-line description. The answer returns probabilities for all three. Your code picks the top option if confidence is high enough, otherwise takes the safe path.

Because many questions share one input, a single call can replace several prompts: a policy check, a topic label, a sentiment score and a language check on the same message. Image and video resolution can be tuned to trade visual detail against token cost.

Integrations are rolling out with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, alongside the google-genai SDKs, according to Google. The Google AI Studio site also hosts a playground and demo apps.

What it means

For developers building agents, the clearest win is the routing and gating layer: tool selection, guardrails and context grading no longer need a generative call that you then have to parse. Calibrated probabilities also make thresholds meaningful, which a model's self-reported "I am 90% sure" in prose is not.

For cost, if the 5x figure holds on your workload, classification-style steps that run on every request become a small line item. Verify with your own traffic: the comparison is against one named model on Google's suite.

For the market, this is a vendor productising a pattern. Smaller specialised heads and classifiers have long done this job; the new part is a general multimodal model offered with that interface and a price per input token only.

The docs are candid about the limits. Choice questions always spread probability across the options, even when none fits, so add a catch-all option, a scope check or a confidence threshold. Score averages can fall between levels, so look at the per-level probabilities and keep scales linear. And "treat user-supplied input as untrusted, since it can argue for its own classification", the docs warn, so keep a human or second check on high-stakes actions.

What we don't know yet

• Independent benchmark results. All performance, latency and cost-comparison claims come from Google.
• How calibration holds on your own data, in other languages and for long multimodal inputs.
• When Gemini in Google Cloud support arrives; Google says only "coming soon".
• Whether the v1beta interface, which supports no conversation history, tool-call blocks, Flex or Priority inference, will change before general availability.
• We found no independent outlet coverage on publication, so everything here is from Google's own pages.

What to watch next

Watch for third-party calibration tests, the Google Cloud launch, and whether OpenAI, Anthropic and open-weight vendors answer with their own score-only endpoints. Developers should also watch the free tier terms if the input contains private data.

Sources

• Google: Gemini Decisions API announcement
• Google AI Studio: Decisions API documentation
• Google AI Studio: Gemini API pricing