Google Launches Gemini Decisions API: No Text Generation, Only Calibrated Probabilities, At Just $0.1 Per Million Tokens
Google has launched the Gemini Decisions API built on the Gemini Decision 1 model. It returns calibrated probabilities for up to 1,000 yes/no, multiple-choice, or scoring questions targeting inputs including text, images, audio, video and PDFs, at a price of $0.10 per million input tokens. This marks a pivotal shift of large language models from content generation to decision-assisting use cases.
Google rolled out a distinct large model product today: the Gemini Decisions API. It is not built for chat, nor does it generate paragraphs of text. You feed it a piece of text, an image, an audio clip, a video, or a PDF, then submit up to 1,000 questions that can be set as yes/no, single-choice or scoring rubric formats, and it returns a calibrated probability for each question.
The so-called calibrated probability, in simple terms, means the model's confidence in its own judgment is highly reliable. For example, if it says there is a 70% probability of a "yes" result, roughly 70% of all similar cases in a large dataset will indeed turn out to be "yes". This feature is critical for scenarios requiring automated decision-making, such as content moderation, work order classification, questionnaire scoring and compliance checks. Traditional large language models give you a natural language response that requires extra parsing work, while this API directly outputs structured probabilities, eliminating the need for follow-up post-processing.
The price is $0.10 per million input tokens. This figure is well worth attention. GPT-4 tier models usually charge several dollars to tens of dollars per million output tokens, but here the cost of judgment tasks is pushed down to near zero. Google clearly wants to make "model-powered judgment" as cheap as calling a regular traditional classifier.
As reported by Michal Sutter, author of MarkTechPost, this API is built on Gemini Decision 1. Multimodal input support means you can directly submit scanned documents, recordings and video frames to ask questions, no pre-conversion to text is required. For example, you can submit a customer service recording and ask "Does the customer express the intention to get a refund?", upload a photo of an invoice and ask "Does the amount exceed $500?", import a PDF of a contract and ask "Does the contract contain an auto-renewal clause?". Up to 1000 questions can be queried in batches in one single request, which is suitable for large-scale evaluation.
However, the actual "calibration" degree of the calibrated probabilities still needs to be tested in practice. Google claims the outputs are calibrated, but performance may vary greatly across different domains, languages and modalities. In addition, the official has not specified whether all 1000 questions are independent, and if there is mutual interference between questions. The price is very low, but if the probabilities are inaccurate, the subsequent cost of manual review will end up even higher.
From the perspective of product form, Google is pushing large language models towards the direction of "decision-making infrastructure". Text generation serves as an entry point, but it is the judgment capability that gets truly embedded into business processes. The Gemini Decisions API is more like a programmable probability switch, and it is not designed to be a chatbot. For developers, this means you can use the same model to handle classification, scoring and filtering tasks, no need to maintain multiple sets of small custom models. For ordinary users, you will not perceive its existence, but it may determine whether your appeal is automatically approved, or whether your content gets restricted reach.
Original link: https://www.marktechpost.com/2026/11/11/google-gemini-decisions-api-gemini-decision-1/
发布时间: 2026-11-12 05:21