Google launches the Gemini Decisions API for fast closed-set calls

The announcement caps that shared input at 256K tokens[1]. The Decisions API guide, which the ai.google.dev docs path now redirects to, sets the context window at 262,144 tokens across the input and every question. Questions run in parallel against that input. They cannot see one another, so adding or removing one leaves the other answers unchanged[2].

Paid input is US$0.10 per million tokens, and output is US$0. The shared input is billed once per request, and each question adds the tokens for its own instructions and criteria[1]. The pricing page lists the same rate, treats free-tier input as free of charge, and also counts formatting tokens on each question[3]. Google says there is no variable thinking-token bill and no separate charge for audio transcription or OCR. A no-cost tier in Google AI Studio is for prototyping[1].

Google reports end-to-end latency of 150 to 250 ms, with median model compute of 50 to 100 ms, for contexts up to 16K tokens, staying nearly flat from 1 to 100 questions[1]. On a suite covering agent routing, reasoning, ranking and multimodal understanding, Google says Gemini Decision 1 leads, with the lowest calibration error. Across that suite, Google says Gemini Decision 1 on the Decisions API is nearly five times cheaper than Gemini 3.5 Flash-Lite on the Interactions API[1].

One request can mix three types[1].

• noul is yes/no. It estimates the probability that a condition holds, and you can describe what true and false mean. The guide keeps that yes probability between 0.01 and 0.99, so a threshold above 0.99 never matches[1][2].
• choice picks from 2 to 255 options and returns the winner plus a normalised probability for every option. The guide sorts option keys alphabetically before scoring[1][2].
• score grades an ordered rubric whose levels start at 0. The guide allows 2 to 255 non-empty levels and returns a continuous score, the probability-weighted average of the level indices, with a probability per level[1][2].

Image and video blocks take a resolution of low, medium, high or ultra_high, which sets visual detail and token cost. Ultra high applies to images only. Question instructions are text only[1][2].

Gemini Notebook uses the API as a real-time router. Google says it classifies intent in milliseconds and serves everyday queries with up to 40% lower latency, reserving deep reasoning models for complex tasks. Stitch's rapid canvas experiment uses it to generate playable wireframes of a product flow in seconds and to apply voice or tap edits almost at once. One click hands the screens to the full Stitch experience. CC, from Google Labs, picks an emoji from more than 200 options in a fraction of a second[1]. Google Search uses the API as a safety layer for MCP tools, detecting adversarial behaviour and content against adversarial evals. Google says that check has high efficacy within the latency envelope[1].

It is in Google AI Studio now, with demos named Doorstep Parcels, Speed Sketch and Find the Zoo Keeper[1]. The guide requires SDK 3.2 or later and documents Python, JavaScript and REST on the v1beta endpoint[2]. Integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, alongside the google-genai SDKs. Support for Gemini in Google Cloud is coming soon[1]. The guide leaves out context caching and the Batch API, and it rejects YouTube URLs and gs:// URIs. Media has to be a Files API URI, a public or signed HTTPS URL, or inline bytes[2].

The announcement checks a breakroom photo and a short note against all three question types in one request. The Python below is that example, plus a print of the yes/no probability and its confidence. The model, input blocks and question fields match the guide[1][2].

from google import genai

client = genai.Client(http_options={"api_version": "v1beta"})

result = client.decisions.create(
    model="gemini-decision-1",
    input=[
        {
            "type": "image",
            "uri": "https://storage.googleapis.com/generativeai-downloads/images/spill.jpg",
            "mime_type": "image/jpeg",
        },
        {
            "type": "text",
            "text": "Breakroom counter check before the 9 AM client meeting.",
        },
    ],
    questions={
        "has_spill": {
            "type": "noul",
            "instructions": "Is there an uncleaned liquid spill on the counter?",
        },
        "next_action": {
            "type": "choice",
            "instructions": "Which cleanup action fits the scene?",
            "criteria": {
                "wipe_with_cloth": "Small counter spill with a cloth or towel already nearby.",
                "dispatch_janitorial": "Large floor hazard, broken glass, or active leak.",
                "ready_for_meeting": "Counter is clean and ready.",
            },
        },
        "severity": {
            "type": "score",
            "instructions": "Rate the severity of the counter mess.",
            "criteria": [
                "Clean and ready.",
                "Minor localized spill.",
                "Major spill or safety hazard.",
            ],
        },
    },
)

spill = result.answers["has_spill"]
print(spill.noul, spill.confidence)

The latency band and the cost comparison are Google's measurements on its own suite. Before automating an action, the guide says to set thresholds from labelled examples of your own traffic, weighing a false positive against a false negative[2].