Google launches Gemini Decisions API for closed-set evaluations

Google launched the Gemini Decisions API on November 11, introducing a specialized endpoint powered by its new Gemini Decision 1 model. The model handles closed-set evaluations rather than free-form generation. The service is available through the Gemini API in Google AI Studio, including an API no-cost tier for prototyping.

The endpoint joins Google’s developer-facing Gemini lineup after its Gemini 4 Hydrogen rollout across AI products. A single request can apply as many as 1,000 questions to a shared input containing text, images, audio, video or PDFs, with a stated context limit of 256,000 tokens. Google positions the service for bounded tasks such as model routing, policy checks, tool selection and context grading.

Applications supply the candidate options, which Gemini Decision 1 scores in one forward pass before returning probabilities and confidence values. The API supports three question formats: noul estimates whether a yes-or-no condition holds; choice selects among two to 255 options and returns normalized probabilities for each; and score grades against an ordered rubric that starts at zero, returning a weighted score and probabilities for each level.

Google prices the model at $0.10 per 1 million input tokens, with no charge for output tokens. The company says the shared input is billed once per request, while each question adds its own instruction and criteria tokens. Google reports median model-compute latency of 50 to 100 milliseconds and end-to-end latency of 150 to 250 milliseconds for contexts up to 16,000 tokens. It says latency remains nearly flat from one to 100 questions. Those figures do not establish performance at the API’s advertised limits of 256,000 tokens or 1,000 questions.

Google also says Gemini Decision 1 posted the best result in seven of nine accuracy rows in its launch comparison. Its table gives the model a 92.2% score on the public JevBench split, compared with 90.9% for GPT-6 Luna and 90.0% for Jev 1.14. Jev led the Mind2Web Multimodal and When2Call rows. Google reported a Top-1 expected calibration error of 0.0250, versus 0.0494 for Jev 1.14 and 0.0721 for GPT-6 Luna; lower values indicate closer alignment between confidence and observed accuracy.

On the same company benchmark workload, Google put the total cost at $0.88 for Gemini Decision 1 and $4.14 for Gemini 3.5 Flash-Lite using the Interactions API, the basis for its claim that the new model was nearly five times cheaper. Jev 1.14 cost less in that comparison, at $0.68. The results are company-run rather than a universal performance ranking. A separate Artificial Analysis Decision Index methodology uses six evaluations spanning document decisions and extraction, tool calling, knowledge, reasoning and invariance, so its results are not directly comparable with Google’s launch suite.

Google said Gemini support in Google Cloud was coming soon but gave no launch date.