Google launches Gemini Decision 1 to score yes/no and choice questions

Google on November 11 launched the Gemini Decisions API, a new endpoint for fast fixed-answer decisions powered by a new model called Gemini Decision 1 (model ID gemini-decision-1). The model takes one input made of text, images, audio, video or PDFs and answers up to 1,000 yes/no, multiple-choice or rubric questions about it in a single pass, returning a probability for each answer instead of generated text.

Google prices it at $0.10 per million input tokens, with zero output tokens billed. The launch puts Google in a small but growing group of “decision model” products: developer Simon Willison describes it as Google joining Jev and OpenAI in the category, and Microsoft put its own Decision-1 into public preview in October.

What the Gemini Decisions API does

A decision model answers bounded questions: questions where the developer supplies the allowed answers in advance. Google describes the Decisions API as giving developers “purpose-built primitives” for choices such as tool selection, policy checks, model routing and context grading. In Google’s words, the model is meant to provide “a cost-effective way to route queries, grade content, or check policy rules”.

Each request carries one shared multimodal input of up to 256K tokens along with up to 1,000 independent questions. Developers can combine modalities in that input; Google’s examples include a call recording paired with account metadata, or a scanned PDF insurance claim paired with damage photos. For images and video, the resolution can be tuned to trade visual detail against token cost.

Every answer comes back with a confidence score from 0 to 1, which Google says lets code send ambiguous cases to a fallback or a human reviewer.

Three question types

A single request can mix three question types:

• noul (yes/no): estimates the probability that a yes/no condition holds, with optional descriptions of what counts as true and false.
• choice (multiple choice): selects from 2 to 255 options, returning the winner and normalized probabilities for every candidate.
• score (ordered rubric): grades the input on an ordered rubric starting at level 0, returning a continuous weighted score and the probability of each level.

Google’s sample code checks a photo of a breakroom counter, plus a short text note, against all three types at once: whether there is an uncleaned spill, which cleanup action fits, and how severe the mess is. The input field uses the same Content block format as Google’s Interactions API.

How Gemini Decision 1 differs from a chat model

The core difference is how the answer is produced. A generative model writes its answer token by token. Gemini Decision 1 instead scores the candidate options in a single forward pass, meaning one run through the network, and returns probabilities directly. That design is why no output tokens are billed.

Google gives these performance figures, all from its own measurements:

• Median latency of 50-100 ms of model compute, or 150-250 ms end-to-end, across multimodal inputs and contexts up to 16K tokens.
• Latency that stays “nearly flat” from 1 to 100 questions per request.
• Cost “nearly 5x cheaper” across Google’s shared benchmark suite than running Gemini 3.5 Flash-Lite through the Interactions API.

The latency figures apply to contexts up to 16K tokens; the announcement does not give a latency figure for inputs near the 256K-token maximum.

Spec	Gemini Decision 1 (as stated by Google)
Model ID	gemini-decision-1
Input modalities	Text, images, audio, video, PDFs
Maximum input	256K tokens
Questions per request	Up to 1,000
Question types	Yes/no, multiple choice (2-255 options), ordered rubric
Output	Probabilities and confidence scores (0 to 1)
End-to-end latency	150-250 ms (contexts up to 16K tokens)
Price	$0.10 per 1M input tokens; output not billed
Access	Gemini API in Google AI Studio

Access and how to start

The model is available through the Gemini API in Google AI Studio, using an API key from AI Studio, and Google has published documentation with Python, JavaScript and REST guides. Google’s Python example calls the API through its genai client with the v1beta API version. The announcement does not mention open weights, regional limits or a waitlist.

Google says teams inside the company already use the Decisions API as what it calls “a low-latency control plane”. Two examples from the launch post:

• Gemini Notebook uses it as a real-time router, sending incoming queries to a model tier and thinking level suited to each request. Google says this serves everyday queries with up to 40% lower latency while keeping deep reasoning models for complex tasks.
• Stitch, Google’s design tool, uses it in a new rapid canvas experiment to generate playable wireframes of a product flow in seconds and to apply edits made by voice or tap almost instantly.

How it compares with Jev, OpenAI and Microsoft

Simon Willison, writing on launch day, calls the new API “almost identical to Jev”: it supports the same three question types and even reuses Jev’s term “Noul” for a yes/no question. He notes that Google’s price matches OpenAI’s GPT-6 Luna decision model at 10 cents per million input tokens with free output, and that Jev remains the cheapest of the three at 4.2 cents per million.

The clearest difference is input coverage. According to Willison, Jev handles text (and JSON) only, OpenAI’s decision model handles text and images, and Gemini Decision 1 covers text, images, audio, video and PDFs.

The category is wider than those three. An October analysis by consultancy La Madre describes Microsoft-Decision-1, put into public preview in Microsoft Foundry on October 9, as the third product it had covered in the category, after Cloudflare’s Clef and OpenAI’s Decisions API. Microsoft’s model is built on Qwen3.5-9B, handles yes/no, multiple-choice and rating questions, and is priced at $0.042 per million input tokens with output free. A LinkedIn post by Cyril E. in October reported that OpenAI had made its Decisions API generally available, with Box using it to route documents and send lower-confidence cases to reviewers, and Datadog using it to score the risk of agent-generated shell commands before running them.

Product	Inputs	Price per 1M input tokens
Gemini Decision 1 (Google)	Text, images, audio, video, PDFs	$0.10, output free
GPT-6 Luna decision model (OpenAI)	Text and images	$0.10, output free
Jev	Text (and JSON)	$0.042
Microsoft-Decision-1 (preview)	Not stated in the source	$0.042, output free

Google’s claims versus early independent evidence

Google says that on benchmarks spanning agent routing, reasoning, ranking and multimodal understanding, Gemini Decision 1 “leads across the suite, with the lowest calibration error”. Calibration means the stated probabilities match reality: answers given 0.9 confidence should be right about nine times in ten. The launch text shown does not name the individual benchmarks or the rival models compared, so the claim cannot be checked from the announcement alone.

The first outside test comes from Willison, who passed a video of a stomping robot in Yonkers to the model with four yes/no questions. The model gave about 0.92 probability that the video contains a robot and about 0.92 that it is set in the USA, against about 0.05 for a dragon and about 0.08 for France. Willison observes that Jev and OpenAI both return 1.0 or 0.0 for yes/no questions, while “Gemini insists on being a little less confident than that”. In his output the yes/no answers are labelled with the type “yes_no” rather than “noul”.

That single test shows the model answering correctly on video and producing graded probabilities rather than hard zeros and ones. It does not, on its own, show that those probabilities are well calibrated across real workloads.

What it means for builders, and the limits

For developers running agents, the pitch is to move the many small, bounded decisions out of expensive generative calls. Google frames this as “shipping features simply by defining options rather than training custom classifiers or waiting on generative calls”. Typical candidates are model routing, tool choice, content moderation against a policy, and grading outputs against a rubric. Multimodal input is the practical draw for teams whose decisions depend on calls, scans or footage, which Willison’s comparison suggests the rival text-only and text-and-image products cannot take directly.

Several cautions from the sources apply:

• Calibration needs local checking. Writing about Microsoft’s model, La Madre argues that “calibration is cheap to test and expensive to assume,” and suggests bucketing a few hundred labelled examples from your own workflow by the model’s confidence and checking accuracy in each bucket. It also warns that a model can be well calibrated on common classes and poorly calibrated on rare ones.
• Design the uncertain path. Google’s own guidance is to route ambiguous cases to a fallback or human reviewer. Cyril E. makes a similar point about OpenAI’s API: every decision needs allowed answers, a confidence threshold and a fallback.
• Logs are easier to audit. La Madre notes that a decision model produces a compact record of options, probabilities, threshold and path taken, which is easier to audit than generated reasoning.
• Numbers are vendor-reported. Google’s latency, cost and benchmark figures come from its own evaluation, and the latency figures cover contexts up to 16K tokens.
• Price is not the lowest. Jev and Microsoft-Decision-1 list lower input prices, though with narrower or unstated input coverage.

What to watch next

The next useful signals are independent calibration results on Gemini Decision 1, especially on audio, video and long PDF inputs, where it currently has no direct rival among the products Willison compared. Developers will also want latency figures for inputs well beyond the 16K-token range Google measured, and clarity on how long the API stays on its v1beta version. With Google, OpenAI, Jev, Microsoft and Cloudflare all offering decision models, pricing and modality coverage across the category are likely to keep shifting.