Google Drops Gemini Decision 1 to Score Choices 5x Cheaper Than Flash
Google's new Gemini Decision 1 model scores up to 1,000 multimodal questions in a single forward pass, returning calibrated probabilities instead of text.
- Google launched the Gemini Decisions API powered by new Gemini Decision 1 model.
- Scores up to 1,000 yes/no, choice, or score questions per request.
- Multimodal inputs: text, images, audio, video, and PDFs supported.
- Priced at $0.10 per 1M input tokens with zero output token charges.
- End-to-end latency of 150-250 ms via single forward pass scoring.
- Available now through Gemini API in Google AI Studio with free tier.
Google has released the Gemini Decisions API in AI Studio. Its Gemini Decision 1 model scores options supplied by the caller and returns probabilities with confidence values. The endpoint targets routing, ranking, policy checks, and other workloads that end in a constrained choice.
One pass, 1,000 questions
Gemini Decision 1 evaluates candidates in a single forward pass, avoiding token-by-token decoding. Google reports end-to-end latency of 150 to 250 milliseconds and median model-compute time of 50 to 100 milliseconds for multimodal inputs and contexts of up to 16K tokens.
Each request can contain up to 1,000 questions, allowing applications to reuse shared context and reduce per-decision overhead. Google says this batching contributes to the endpoint’s lower cost.
Requests can mix three question types, so one shared context can support policy checks, route selection, and rubric grading in the same call:
| Question type | Purpose | Output |
|---|---|---|
| yes/no | Test whether a condition holds | Probability and confidence |
| choice | Select among 2 to 255 supplied options | Selected option, probabilities, and confidence |
| score | Grade an input against a rubric | Typed score and confidence |
Confidence becomes a control signal
Every answer includes a confidence value from 0 to 1. Applications can send high-confidence decisions directly into a workflow while routing uncertain cases to a person, a generative model, or another fallback.
Google reports leading results on its benchmark suite for agent routing, reasoning, ranking, and multimodal understanding, along with the lowest calibration error. Calibration measures whether predicted probabilities match observed outcomes. Among decisions assigned a probability of 0.7, a well-calibrated model should be correct about 70% of the time. The results are company-reported and should be validated against each application’s traffic.
The bill moves to inputs
Google prices Gemini Decision 1 at $0.10 per 1 million input tokens and charges nothing for output tokens. A no-cost API tier is available in Google AI Studio for prototypes.
The pricing also excludes separate charges for thinking tokens, audio transcription, and optical character recognition. Typed decisions still come back in the response, but they produce no output-token bill.
In Google’s shared benchmark comparison, Gemini Decision 1 through the Decisions API cost about one-fifth as much as Gemini 3.5 Flash-Lite through the Interactions API, even when the Flash-Lite calculation assigned zero billed output tokens. Production savings will depend on input size, batch utilization, retries, and fallback rates.
Hard boundaries define the fit
The endpoint returns typed decisions drawn from the question and candidate set supplied by the application. It provides no free-form prose, inspectable reasoning trace, or novel answer outside those bounds. Tasks that require synthesis, explanation, or open-ended generation still need a generative model.
Google’s announcement describes support for text, images, audio, video, and PDFs. The linked BenchLM model card currently lists text, images, and PDFs as inputs, while describing the outputs as typed decisions with probabilities. Model parameters remain undisclosed. Developers planning to use audio or video should verify support in the current SDK and API documentation.
A small SDK surface
Google’s Python client exposes the endpoint through client.decisions.create(). The following sketch identifies the client method and model name; request fields should follow the schema in the installed SDK version:
from google import genai
client = genai.Client()
result = client.decisions.create(
model="gemini-decision-1",
# Add questions, candidates, and input content here.
)Three bundled AI Studio apps demonstrate the interaction pattern: Doorstep Parcels performs image question answering, Speed Sketch judges live drawings, and Find the Zoo Keeper handles visual search.
Decision endpoints become a category
Bounded-decision APIs formalize a pattern developers have often implemented with constrained JSON, one-token responses, or custom classifier heads. OpenAI put its Decisions API into public beta on October 6, 2026. Microsoft followed on October 9, 2026, with Microsoft-Decision-1. Google’s stated differentiators are its 16K-token context window and multimodal input support.
Where it fits in a stack
Teams running agents, moderation systems, retrieval pipelines, or model gateways can apply the endpoint to several recurring tasks:
- selecting a tool or model for an agent request
- checking content against a policy
- scoring retrieved passages for relevance
- ranking supplied candidates
- triaging multimodal inputs for review
A production evaluation should measure calibration, task accuracy, tail latency, and total cost on representative traffic. Thresholds also need explicit fallback behavior for uncertain decisions and inputs that fall outside the supplied candidate set.
Comments
