Google added a decision model API that scores 1,000 choices at once

A team that routes agent requests or grades an eval pays for a long text answer. All it wants is yes or no. Google launched the Gemini Decisions API today. The new model, Gemini Decision 1, scores a candidate answer in one pass. It returns a probability, not text.

Google says the API supports three question shapes. Yes-no. Multiple-choice with 2 to 255 options. Rubric scoring. One request can carry up to 1,000 independent questions. The input can be text, an image, audio, video or a PDF, up to 256 thousand tokens.

Google says the model's median compute time is 50 to 100 milliseconds. The end-to-end response is 150 to 250 milliseconds. Google lists the price at 10 cents per million input tokens. Output tokens are free.

The price next to its rivals

Google says the model is near five times cheaper than Gemini 2.5 Flash-Lite. That figure is from Google's own test on comparable decision work.

OpenAI's decision model charges the same 10 cents per million input tokens, Simon Willison notes. TypeSafe's Jev is cheaper. It charges 4.2 cents per million. Jev was the first decision model on the market. Reveneau uses it to grade its own eval suite.

What is new in it

The three decision shapes match Jev and OpenAI. The new part is the input types, Willison writes. Jev reads text and JSON. OpenAI reads text and images. Gemini Decision 1 is the first that also reads audio, video and PDFs. That widens the jobs an engineer can hand to one call.

The confidence scores are different too. Willison showed the model returning a 0.923 probability on a yes-no question. The two earlier models hand back 1.0 or 0.0. A team that routes requests can treat that number as a threshold. A routing rule becomes a measurable choice.

Google says Gemini Decision 1 reaches "the lowest calibration error" on its own tests. The tests cover agent routing, reasoning and multimodal work. The claim is Google's. The tests are Google's.

The product managers behind it

The post names Min Choi as the product manager and Felix Yu as the principal scientist. The product is available in Google AI Studio. There is a free tier for building. Google says ecosystem support is rolling out in OpenRouter, LangChain and Vercel.

Three vendors, three trade-offs

The three vendors now cover the decision slot with different trade-offs. Jev is the cheapest. It has the longest track record in production grading. OpenAI matches Google on price. It reads text and images. Gemini Decision 1 reads every file type. It is backed by Google's model catalogue.

A team running a grader or a router has a reason to test all three on its own cases. Read the confidence scores side by side. Each vendor calibrates its scores against its own tests. The same number from two models will mean two different things.

Source

Introducing the Gemini Decisions API, Google, by Min Choi and Felix Yu. Technical notes: Gemini Decisions API, Simon Willison.