Google launches Gemini Decisions API for batch decisions
Google’s new Gemini Decision 1 model handles up to 1,000 multimodal choices in one pass, targeting fast, low-cost routing and policy checks.
What happened
Google introduced the Gemini Decisions API, a new endpoint built around Gemini Decision 1 for bounded, repeatable choices rather than open-ended text generation. Developers can submit one text, image, audio, video or PDF input and ask up to 1,000 yes/no, multiple-choice or rubric-scored questions in a single request.
Google says the model returns probabilities and confidence scores with end-to-end latency of roughly 150–250 milliseconds. Pricing is listed at $0.10 per million input tokens, with no output-token charge. The service is available through the Gemini API in Google AI Studio.
Why the design matters
The product targets a layer of AI infrastructure that is usually hidden inside applications: deciding which tool to call, whether a document satisfies a policy, how to route a request between models, or whether retrieved context is relevant. Those tasks often do not need a conversational answer, but they do need consistent scoring and predictable latency.
The batch format is especially notable. A company screening a large set of documents, images or video segments can evaluate many criteria against the same multimodal context without repeatedly invoking a generative model. That could reduce both latency and orchestration complexity, although Google’s public announcement does not yet establish how accuracy compares with dedicated classifiers or human review.
What remains unproven
The API is optimized for closed-set decisions, so it is not a general replacement for reasoning models. Confidence scores also require calibration evidence across domains; a numerical probability is useful only when it tracks real-world error rates. Google’s examples show the intended workflow, but independent testing will determine whether the price and speed survive production workloads.
For developers building agents, the release signals that model providers are packaging decision-making as a specialized primitive rather than treating every step as chat completion.