Google introduced the Gemini Decisions API on November 11, a new endpoint for fast, closed-set decisions powered by its new Gemini Decision 1 model. The Gemini Decisions API evaluates up to 1,000 yes-or-no, choice and rubric-scoring questions in a single pass across text, images, audio, video and PDFs. It is available through the Gemini API in Google AI Studio. The announcement came in a blog post by Min Choi, a product manager at Google DeepMind, and Felix Yu, a principal scientist there. Google chief executive Sundar Pichai also posted about the launch on X the same day.
Gemini Decisions API: how the scoring works
Developers send one multimodal input of up to 256,000 tokens along with as many as 1,000 independent questions. Rather than generating answer text token by token, Gemini Decision 1 scores the candidate options in a single forward pass and returns probabilities and confidence scores. Questions come in three forms. A yes-or-no question returns the probability of each answer. A choice question picks from 2 to 255 options and returns a normalized probability for every candidate. A score question applies an ordered rubric and returns a continuous weighted value. Every answer carries a confidence score from 0 to 1, so code can send unclear cases to a fallback or a human reviewer. Google positions the Gemini Decisions API for tool selection, policy checks, model routing and context grading.
Gemini Decisions API speed and price
Google said the service returns probabilities and confidence scores in 150 to 250 milliseconds end to end, with 50 to 100 milliseconds of median model compute. The company said latency stays nearly flat from 1 to 100 questions per request across multimodal inputs and contexts up to 16,000 tokens. Gemini Decision 1 is priced at $0.10 per 1 million input tokens, with no charge for output tokens because the model produces none. A no-cost tier is available in Google AI Studio for prototyping. Google said the Gemini Decisions API is nearly five times cheaper across its shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API. The input can combine modes, for example a call recording with account metadata, or a scanned PDF claim with photos of the damage.
Benchmark results come from Google
On benchmarks spanning agent routing, reasoning, ranking and multimodal understanding, Google said Gemini Decision 1 leads across the suite, with the lowest calibration error. The company reported a score of 92.2 percent on JevBench, against 90.9 percent for GPT-6 Luna and 90.0 percent for TypeSafe Jev 1.14. Its Top-1 expected calibration error was 0.0250, against 0.0721 for GPT-6 Luna and 0.0494 for TypeSafe Jev 1.14. Those figures are Google’s own, and no outside testing has been cited. The Gemini Decisions API is built on Gemini’s native multimodal architecture, so it can process text, images, audio, video and PDFs directly in a single forward pass, and questions in one request cannot see each other.
Google built the Gemini Decisions API for work where the answer is one of a set list, not a paragraph of prose. A policy check returns yes or no with a probability. A routing step picks one tool from a list and reports how sure the model is about each option. A grading step scores a response against a rubric and returns a single number. Because the Gemini Decisions API returns probabilities instead of text, developers can set a threshold and send low-confidence cases to a person or a second model. That fits systems that run thousands of small judgments a day, such as sorting support tickets, checking claims against policy, or grading model outputs before they reach a user.
The launch gives developers a narrow tool for a common job: asking a model to choose between set options rather than write open-ended text. A yes-or-no check on a document, a pick from a list of tools, or a grade against a rubric can now run as one scored request with a confidence value attached. Google has published a guide and a model page for the service, and the no-cost tier lets developers test it before paying for input tokens.