📰 Key Takeaways

Google has launched the Gemini Decisions API, a new endpoint designed for “closed-option decisions,” powered by the new Gemini Decision 1 model (gemini-decision-1) and available through the Gemini API in Google AI Studio. It fits scenarios where you’re picking from a limited set of options, like tool selection, policy checks, model routing, and contextual scoring. The core mechanism: instead of generating answer text token by token, it scores the candidate options directly in a single forward pass and returns probabilities along with confidence scores. Google lists end-to-end latency at roughly 150 to 250 milliseconds (median model compute of 50 to 100 milliseconds), with input priced at $0.10 per million tokens and no charge for output tokens. Google says latency stays nearly flat from 1 to 100 questions, that overall cost is nearly 5x cheaper than Gemini 3.5 Flash-Lite paired with the Interactions API, and that it leads benchmarks in agent routing, reasoning, ranking, and multimodal understanding, with the lowest calibration error.

Inputs can mix text, images, audio, video, and PDFs, up to 256K tokens, and a single request can include up to 1,000 independent questions. You can also adjust image and video resolution to balance detail against token cost. For example, you could evaluate a call recording together with account data, or a scanned claim PDF together with photos of the damage. There are three question types: yes/no judgments estimate the probability that a condition holds, and you can attach true/false criteria; choice questions pick a winner from 2 to 255 options and return a normalized probability for each candidate; score questions rate on an ordinal scale starting from 0, returning a continuous weighted score and the probability of each level. Every answer comes with a confidence score from 0 to 1, which your code can use to hand ambiguous cases to a fallback model or human review. Input fields reuse the Content block format from the Interactions API. See the original post for more details.

💬 JudyAI Lab Take

Google’s Gemini Decisions API turns “closed-option decisions” into their own endpoint. Instead of generating text token by token, it scores the candidate options directly in a single forward pass. Google lists end-to-end latency at roughly 150 to 250 milliseconds, and this design direction is worth a look for AI builders.

For things like tool selection, policy checks, and model routing, where there are only a handful of options, a lot of people have been tossing the question at a general-purpose generative model and then parsing the answer out of the output text. The thinking behind this news is: when the problem is really “pick one out of N,” you should use an interface built for exactly that. From the original post, input costs $0.10 per million tokens and output is free, Google says it’s nearly 5x cheaper than Gemini 3.5 Flash-Lite paired with the Interactions API, a single request can carry up to 1,000 questions, and latency stays nearly flat from 1 to 100 questions. Another detail worth noting is that every answer comes with a confidence score from 0 to 1, so your code can route ambiguous cases to a fallback model or human review. That turns a “decision” from an uncontrollable blob of text into a structured signal you can set thresholds on and route. Of course, all of these numbers and benchmark results are Google’s own claims, so you’ll still want to validate actual performance on your own data.

A good first step: take inventory of your AI workflows, find the steps that are really just routing or classification, and try setting a confidence threshold for them, sending anything below the threshold to human review.

📅 Source Info

🔗 Further Reading