Google is introducing the Gemini Decisions API. This is a new API interface for making quick decisions from a predefined set of options. The API is powered by the new Gemini Decision 1 model.
The Decisions API offers specially developed features for well-defined choices, such as selecting tools, checking policy rules, routing requests to models, and assessing context. Users can submit a single multimodal input consisting of text, images, audio, video, and PDFs, with a maximum of 256,000 tokens. In addition, up to 1,000 independent questions can be asked.
Instead of generating answers token by token, Gemini Decision 1 evaluates possible answers in a single processing step. The model returns probabilities and confidence scores, with a total response time of 150 to 250 milliseconds. The cost is $0.10 per million input tokens. No charges are applied for output tokens.
According to Google, Gemini Decision 1 outperforms competing models on a series of benchmarks for AI agent routing, reasoning, ranking, and multimodal understanding. Additionally, the model exhibits the smallest calibration error. This means that the reported probabilities correspond relatively well with the actual probability of a correct answer.
Because no output tokens are charged, Google states that Gemini Decision 1 with the Decisions API is nearly five times cheaper than Gemini 3.5 Flash-Lite with the Interactions API on the benchmarks used.
Multimodal evaluation of three types of decisions
Gemini Decision 1 is based on Gemini’s multimodal architecture. Different types of information can be combined into a single shared input. Consider, for example, a recording of a phone call combined with customer data. Or a scanned damage report with photos of the damage. For images and videos, the resolution can be adjusted. This allows you to control the amount of visual detail and token consumption.
With the Gemini Decisions API, Google is entering the growing market for decision models, according to SiliconANGLE. Startup TypeSafe AI introduced the Jev model back in September, while OpenAI launched a similar API last month.