SI Agent News
DevelopingGeminiPolicyResearch

Google launches Gemini Decisions API for structured multimodal scoring

The new API allows developers to extract bounded judgments like yes/no answers and rubric scores with confidence estimates across text, images, and video.

150–250 millisecondsEnd-to-end latency for 16K token contexts

The short version

Why it matters

This tool helps move AI from 'chatting' to 'deciding' by providing structured data that software can actually use to trigger actions. It allows for more rigorous evaluation of AI outputs by providing confidence scores alongside the results.

Structured outputs for developers

Google launched the Gemini Decisions API and the Gemini Decision 1 model on November 11, 2026. This new tool is designed for developers who need models to perform specific tasks like policy checks, tool selection, and context grading. Unlike standard models that generate conversational text, this API focuses on returning bounded judgments.

The model is built to handle multimodal inputs, meaning it can process text, images, audio, video, and PDFs. It is currently available through the Gemini API in Google AI Studio. Google has also indicated that support for Gemini in Google Cloud is coming soon.

By focusing on structured outputs, the API helps researchers and developers build more reliable workflows. It allows a system to evaluate a dataset or a document and provide a clear, quantifiable result that a computer program can interpret easily.

How the three question types work

The Gemini Decisions API handles three distinct types of questions, each returning different data types. For yes/no questions, the model returns a probability of 'yes' between 0.01 and 0.99, along with a confidence estimate. This allows developers to set specific thresholds for when a 'yes' is considered certain enough to trigger an action.

The choice type allows a question to evaluate between 2 and 255 different options. The model returns the selected option, the probabilities for every other option, and a confidence score. Because the model allocates probability even when no option fits, Google suggests using a catch-all option or a separate scope check to ensure accuracy.

The score type uses a rubric with 2 to 255 ordered levels to evaluate an input. Instead of just giving an integer grade, it returns a probability-weighted expected score. For example, if a rubric has three levels and the model sees a distribution of probabilities across them, it calculates a continuous score that summarizes that distribution.

Confidence and calibration

One of the key features of the Decisions API is the inclusion of confidence estimates. Google defines confidence as an estimated probability from 0 to 1 that the top answer provided by the model is correct. For yes/no questions, the model calculates this by taking the larger of the yes and no probabilities.

It is important for users to understand that a high confidence score is an estimate and not a guarantee of correctness. A model can be very confident in an answer that is actually wrong. Google advises developers to set their own thresholds by testing the model against their own specific traffic and weighing the risks of false positives against false negatives.

To help developers understand these estimates, the documentation references calibration curves in machine learning. This concept helps users evaluate how well the model's predicted probabilities align with actual outcomes. Google claims the model has the lowest calibration error across its benchmark suite, though these are vendor-reported results.

Technical limits and integration

The API allows for 1 to 1,000 questions to be run independently against a shared input within a 262,144-token context window. However, the questions cannot see one another during the process. This means the application code must be designed to enforce the logic between different checks, such as ensuring a routing answer does not override a failed documentation check.

Several features are currently unsupported, including the Batch API, caching, and function calling. The model is currently accessible via the v1beta version of the Gemini Developer API. Developers can use the free tier, which allows for 200 requests per minute and 10,000 requests per day, though these limits are not guaranteed by Google.

The API is designed for speed, with tests showing nearly flat latency across 1 to 100 questions. Even with large contexts of up to 16K tokens, the end-to-end latency remains between 150 and 250 milliseconds. This makes it suitable for applications that require quick, structured decisions in real-time workflows.

Practical research applications

While Google has not reported a specific university deployment, they provided an illustrative example of how a research office might use the tool. A university repository team could submit a dataset description and a PDF to the model to check for required documentation. They could then ask the model to choose a review queue and score the completeness of that documentation.

For a research office documenting these evaluations, the API provides a way to create an AI evaluation card. This card can record the specific rubric and the escalation criteria alongside the project's broader model evaluation suite. This creates a transparent record of how the AI reached its specific scores and judgments.

Because the score is a continuous value rather than a simple integer, it provides a more nuanced view of the model's reasoning. A score summarizing a distribution across rubric levels does not specify a single workflow action on its own. Developers must interpret the probability-weighted score to decide how to proceed with the next step in their specific pipeline.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by CASRAI · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news