Gemini Decisions API

Google has become the third company to sell a decision model, a kind of API that answers closed questions about an input with numbers instead of written text. A developer sends one input, such as a document or a video, along with up to 1,000 questions, each a yes/no question, a multiple choice or a score on a rubric. In Google's announcement, also covered here, the company says Gemini Decision 1 scores the options in a single pass rather than generating an answer word by word, which makes it suited to quick routing and filtering checks. Simon Willison, who explained the category when TypeSafe AI released its Jev model in September, writes that Google's API is almost identical to Jev, down to reusing Jev's name, Noul, for a yes/no question.

Google matched OpenAI on price rather than Jev. Willison puts Gemini Decision 1 at 10 cents per million input tokens with free output, the same as OpenAI's GPT-6 Luna decision model, for which he built a plugin in October, while Jev stays cheapest at 4.2 cents. The difference is in what each will read. Jev takes only text and JSON and OpenAI's takes text and images, while Google's also accepts audio, video and PDFs.

Willison tested the video input with four yes/no questions about a clip of a stomping robot in Yonkers. Gemini put the chance of a robot at about 0.92 and of a US setting at about 0.92, against roughly 0.05 for a dragon and 0.08 for France. He notes that Jev and OpenAI return a flat 1.0 or 0.0 for yes/no questions, while Gemini stops short of certainty. That appears to be deliberate: Google's documentation describes the yes/no output as a calibrated probability held between 0.01 and 0.99, and suggests developers pick their cut-off points from labeled examples of their own traffic. A graded answer gives code something to act on, since Google says ambiguous cases can be sent to a fallback or a human reviewer. Willison's four questions show the format, not whether those probabilities are accurate.