Gemini Decisions API: Price, Speed, When to Use It

Gemini Decision 1 (gemini-decision-1) is Google’s closed-set decision model powering the Gemini Decisions API, launched on November 11, 2026. It evaluates up to 1,000 typed questions against shared multimodal input in one forward pass, returning calibrated probabilities in 150 to 250 ms end-to-end for $0.10 per 1 million input tokens with zero output tokens.

On November 11, 2026, Google DeepMind Product Manager Min Choi and Principal Scientist Felix Yu introduced the Gemini Decisions API on the official Google blog, alongside an announcement from Google CEO Sundar Pichai on X. Rather than generating text for humans to read and code to parse, the endpoint scores closed sets of options and returns probabilities that software can branch, sort, and route on directly.

What Is Gemini Decision 1 and How Does the Gemini Decisions API Work?

Gemini Decision 1 is a multimodal decision model (gemini-decision-1) available on the Gemini Developer API (v1beta) in Google AI Studio. As the official Decisions API documentation states, generative models produce text for people to read, whereas generating and parsing text when code only needs a judgment to branch on adds latency and brittle parsing logic.

You call POST https://generativelanguage.googleapis.com/v1beta/decisions (with x-goog-api-key) or client.decisions.create(model="gemini-decision-1", input=..., questions=...) using SDK 3.2 or later (pip install -U google-genai or npm install @google/genai@latest). The model scores the questions map (an object of named question keys) in one forward pass, emits zero output tokens, and returns calibrated probabilities and confidence values.

Questions are scored in parallel and cannot see one another; adding or removing a question never alters other answers. Because the model never sees question keys ([A-Za-z0-9_-]+, 1 to 64 ASCII characters), the full prompt must appear inside each question’s text-only instructions field. Per the Gemini Decision 1 model page, text generation, thinking, function calling, structured outputs, Batch API, caching, Flex and Priority inference, Live API, computer use, code execution, file search, grounding, URL context, and image or audio generation are all unsupported.

What Are the Three Question Types in the Gemini Decisions API?

The Gemini Decisions API supports three question types (noul, choice, and score) that can be mixed in one request of up to 1,000 questions. Every response includes confidence (0 to 1 probability that the top answer is right).

1. Yes/No (`noul`)

The noul type returns the calibrated probability that a binary condition holds, with optional true_description and false_description criteria. Probabilities always fall between 0.01 and 0.99 (a threshold above 0.99 never matches). The probability of no is 1 - noul, and confidence equals max(noul, 1 - noul).

2. Multiple Choice (`choice`)

The choice type evaluates 2 to 255 options mapped in criteria (option IDs of 1 to 128 characters, up to 512 UTF-8 bytes; empty description strings allowed when IDs are self-explanatory). It returns the winner, normalized probabilities summing to 1, and confidence equal to the top outcome’s probability. Before scoring, the API sorts option keys alphabetically in UTF-8 byte order and breaks ties to the alphabetically first key. Avoid positional keys like none_of_the_above; use other instead. Because probability is always distributed across the given options even when none fits, include a catch-all option or a noul check when needed.

3. Ordered Rubrics (`score`)

The score type grades input against 2 to 255 ordered levels (2 to 10 recommended), indexed from 0, using an array of non-empty criteria descriptions from lowest to highest. It returns a continuous probability-weighted score, per-level probabilities, a legend, and confidence. Because different distributions can share the same expected value, check per-level probabilities and confidence alongside the score. Pick the highest-probability entry for a discrete level, divide by the highest level index to normalize to 0 to 1, and keep scales linear so array-index averaging is not skewed. See our guide to LLM-as-a-judge evaluation for more on rubric design.

How Much Does Gemini Decision 1 Cost? (Decisions API Pricing)

Gemini Decision 1 costs $0.10 per 1 million input tokens on the paid tier as of November 12, 2026, and output tokens are not billed (output token usage is always 0). Google also offers a free prototyping tier in Google AI Studio. Understanding what a token is in AI shows why skipping output billing changes classification economics.

Per the official Gemini pricing page, shared input is billed once per request across up to 1,000 questions; each question adds only its instruction, criteria, and formatting tokens. Text, image, video, audio, and PDF inputs all cost $0.10 per 1M tokens (PDF pages and video frames count at the image token rate), with no output, thinking, OCR, or audio transcription fees. Context caching and the Batch API are unavailable, so repeated input across separate requests is billed in full. Free-tier inputs are used to improve Google products; paid-tier inputs are not.

In its launch blog, Google states that with zero output tokens billed, Gemini 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. That is a Google-reported benchmark result, not a simple list-price ratio. For context from the Gemini 3.5 Flash-Lite pricing table as of November 12, 2026, gemini-3.5-flash-lite lists at $0.30 per 1M input tokens and $2.50 per 1M output tokens (including thinking tokens). By list-price arithmetic, 10,000 input tokens cost $0.001 on Decision 1 versus $0.003 on Flash-Lite before output tokens. Compare broader rates in our AI API pricing comparison and how to reduce AI API costs.

How Fast Is the Gemini Decisions API and What Are Its Technical Limits?

The Gemini Decisions API delivers an official median latency of 50 to 100 ms of model compute (150 to 250 ms end-to-end). Google states that latency stays nearly flat across multimodal inputs and contexts up to 16K tokens, from 1 to 100 questions per request (do not assume flat latency up to 1,000 questions or 262,144 tokens).

Hard limits as of November 12, 2026 include 1 to 1,000 questions per request, a 262,144-token input window (the 256K-class window covering shared input plus all questions), 0 output tokens, and 100 JSON nesting levels. The shared input accepts text, images (png, jpeg, webp, heic, heif, gif, bmp, tiff), audio (wav, mp3, mpeg, aiff, aac, ogg, flac, m4a, opus, webm, l16, alaw, mulaw), video (mp4, mpeg, mov, avi, webm, wmv, x-flv, 3gpp), and PDFs (application/pdf only), with low, medium, high, and image-only ultra_high resolution settings. Inline base64 allows 100 MB per request (50 MB for PDFs), and Files API or HTTPS URLs allow 15 MB per external URL (no YouTube or gs:// URIs).

How Does Gemini Decision 1 Perform on Benchmarks?

Benchmark data comes from two separate suites as of November 12, 2026. First, Google reports in its launch post that on its internal benchmarks spanning agent routing, reasoning, ranking, and multimodal understanding, Gemini Decision 1 leads across the suite, with the lowest calibration error.

Second, independent firm Artificial Analysis scored Gemini Decision 1 at 66 on its Artificial Analysis Decision Index v2.1, in line with Base Labs’ Elephant 9B and TypeSafe’s Jev 1.14, and 7 points behind leaders Nace AI’s Drex Max v1.1 and Fastino’s GLiDE v1.2 at 73. Artificial Analysis reported that Gemini Decision 1 leads knowledge and reasoning tests (AA-Omniscience-Decision and HLE-MultiChoice) while trailing on other tests, including AA-Paperclip document classification and extraction. The two results reflect different evaluation suites.

When Should You Use the Decisions API vs the Interactions API?

Use the Gemini Decisions API when your code needs calibrated probabilities over closed options; use the Interactions API when you need generated tokens. See our guide on how to choose an AI model for routing frameworks.

Dimension / Workload	Gemini Decisions API (`gemini-decision-1`)	Gemini Interactions API ( e.g., `gemini-3.5-flash-lite`)
Paid Price (as of Nov 12, 2026)	$0.10 / 1M input tokens; output not applicable (free tier in AI Studio)	$0.30 / 1M input; $2.50 / 1M output (includes thinking)
Best Use Cases	Tool selection, policy checks, model routing, context grading, real-time UI/audio/video reaction	Generated text, summaries, explanations, custom JSON schemas, tool calls with arguments, multi-step reasoning
Caching & Batch API	Not supported (shared input billed once per request)	Supported on standard generative models

Google’s blog cites four first-party uses: Gemini Notebook routes queries to cut everyday latency by up to 40%; Google Search screens MCP tools against adversarial behavior; Stitch powers rapid wireframes and voice/tap edits; and CC from Google Labs selects from 200+ emojis in a fraction of a second. Developers can try the Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper demos or the AI Studio Decisions Playground. Ecosystem integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode alongside google-genai, with support coming soon to Gemini in Google Cloud.

FAQ

Is the Gemini Decisions API available on Google Cloud Vertex AI?

As of November 12, 2026, the Gemini Decisions API is available only on the Gemini Developer API (v1beta) in Google AI Studio with gemini-decision-1. Google states that support is coming soon to Gemini in Google Cloud, so Vertex AI general availability has not launched yet.

What happens when safety filters block a Gemini Decisions API request?

When input safety filters block a request, the API returns HTTP 400 instead of answers, with error codes such as safety, prohibited_content, blocklist, image_safety, or content_blocked. Google advises treating user input as untrusted, writing explicit criteria, testing adversarial inputs, and keeping a second check on high-stakes actions like payouts.

How does shared input billing work across multiple questions?

Within a single request, the shared input is billed once at $0.10 per 1 million tokens on the paid tier across up to 1,000 questions. Each entry in the questions map adds only its own instruction, criteria, and formatting tokens, and output tokens are not billed.

Can Gemini Decision 1 return explanations or tool-call arguments?

No. Gemini Decision 1 has an output token limit of zero and does not support structured outputs, function calling, or thinking tokens. It can choose a tool ID from a closed choice list, whereas generating arguments, explanations, or custom JSON objects requires the Interactions API.

Gemini Decision 1 and the Decisions API: Price, Speed, and When to Use It

What Is Gemini Decision 1 and How Does the Gemini Decisions API Work?