Gemini Decisions API
Google has released a model called Gemini Decision 1, joining Jev and OpenAI in what commentators are calling the decision-model category. The launch was surfaced by Simon Willison, who notes that the new API closely mirrors Jev's design: it accepts the same three question types, namely yes/no, multiple choice and rubric scoring, and even reuses Jev's term "Noul" for a yes/no question. That kind of interface compatibility lowers the cost of switching between providers, since a developer can point existing question-and-answer logic at a new endpoint without redesigning the request format.
Pricing puts Gemini Decision 1 at 10 cents per million input tokens, with output tokens free. That matches OpenAI's GPT-6 Luna decision model exactly. Jev remains the cheapest of the three at 4.2 cents per million. The three products therefore split on two axes: price and what media they can read. Jev is text and JSON only, OpenAI's offering handles text and images, and Gemini Decision 1 is described as fully multimodal, covering text, images, audio, video and PDFs.
Willison tested the video capability by submitting a file he calls the Yonkers stomping robot video alongside four yes/no questions: whether the clip contains a dragon, whether it contains a robot, whether it is set in France, and whether it is set in the USA. Jev and OpenAI return hard 1.0 or 0.0 for yes/no questions. Gemini instead returns a probability for each, in a JSON structure that pairs each question's identifier with a type of "yes_no" and a floating-point score.
In that run the dragon question came back at roughly 0.0498, the robot question at about 0.9234, France at about 0.0833 and the USA at about 0.9230. Read plainly, the model is fairly confident the video contains a robot and is set in the USA, and fairly confident it contains no dragon and is not set in France. A developer would typically apply a threshold, such as 0.5, to convert these into booleans.
This suggests the graduations are the more interesting design choice. A continuous probability lets callers express uncertainty, set their own cutoffs, and route borderline cases to a human or a second model, which is awkward when a provider only hands back a bare 1 or 0. The trade-off is that consumers of the API must now own thresholding logic and decide how to treat mid-range scores, and teams comparing providers need to remember that a 0.92 from Gemini is not directly comparable to a hard 1.0 from Jev or OpenAI.
Why it matters: teams building classifiers, moderation filters or content taggers now have a third decision-model option, and the multimodal coverage is the differentiator, since video, audio and PDF inputs can be judged without a separate transcription or frame-extraction pipeline. Inference: the matching price with OpenAI and the reuse of Jev's request vocabulary point to a commodity market where media support and output format, not cost, become the deciding factors.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Gemini Decisions API Google join Jev and OpenAI in the decision model category with "Gemini Decision 1". This new model API is almost identical to Jev: it supports the same three types of questions (yes/no, choice, and rubric-scoring) and even reuses Jev's term "Noul" for a yes/no question. It's priced the same as the GPT-6 Luna decision model: 10 cents per million input tokens, and output is free. Jev remains the cheapest of the three, at 4.2 cents per million. Jev is text (and JSON) only. OpenAI can handle text a