Big models are great at writing long answers, but all a lot of businesses really need is a “yes” or “no.”
Google launched the Gemini Decisions API on November 11, behind a dedicated new model, Gemini Decision 1. It is not responsible for writing articles, but only for making closed judgments: true or false questions, multiple choice questions, and scoring questions.
The usage is as follows: you upload a material, which can be text, image, audio, video or PDF, up to 256K tokens, and then throw up to 1000 independent questions. The model only runs a forward calculation once and gives a probability and confidence level to each option.
In the past, when doing the same thing, we either split it into hundreds of calls, or asked the large model to write a paragraph and then parse the answer from it. Now done in one go, end-to-end latency is about 150 to 250 milliseconds.
The price is its most ruthless aspect: $0.10 per million tokens for input and no charge for output tokens. Google says it’s nearly 5 times cheaper than a solution put together using Gemini 3.5 Flash-Lite. It is now available on Google AI Studio, and you can try it with a free quota. The gateway tool LiteLLM also already supports it through the /v1/decisions interface.
Suitable scenarios: Which tool should the agent adjust, whether a conversation violates policies, which model should the request be routed to, and whether the quality of the retrieved context is sufficient. You can also feed a customer service recording and account information together to check more than a dozen items at once.
Its limitations are also obvious, and it does not care about open-ended questions. But in the agent pipeline, a large number of nodes are inherently judgmental questions. Separating these nodes from expensive large models can save a lot on bills.