What the Gemini Decisions API Actually Does
Picture a claims adjuster staring at ten thousand damage reports, each needing the same twenty checks before a human ever sees it. Google’s Gemini Decisions API, announced on November 11, was built to answer questions like those in bulk — fast, cheap and with a confidence score attached.
The new endpoint is powered by a model called Gemini Decision 1, and it is available now through the Gemini API in Google AI Studio. It does not write essays, poems or code. It judges.
Send it one piece of material — a document, a photo, an audio clip, a video or a PDF — along with up to 1,000 independent questions about it. In a single pass, it scores every question and returns an answer with a probability for each one.
That is a different kind of AI from the chatbots most people know. A chatbot answers one question at a time, in sentences. This API answers a thousand at once, in numbers.
How One Pass Replaces a Thousand Answers
Ordinary generative models produce text token by token, like someone typing. Each word depends on the one before it, so a long answer takes a long time. The Decisions API skips that step entirely.
Instead, Gemini Decision 1 reads the input once and scores the candidate answers in a single forward pass. The result, according to Google, is end-to-end latency of 150 to 250 milliseconds — roughly the blink of an eye.
That speed barely changes as the workload grows. Google says latency stays nearly flat as a request goes from one question to a hundred, at contexts up to 16,000 tokens. A thousand-question checklist costs almost the same time as a one-question check.
The shared input can run up to 256,000 tokens, covering the evidence plus the instructions and choices for every question. Developers can mix formats in a single request: text, images, audio, video and PDFs.
Three Kinds of Questions, One Request
The API supports three decision types, and each one returns a confidence score — an estimated probability, from zero to one, that the top answer is correct. That number is the whole point.
The first type is a yes-or-no check. It returns the probability that a condition holds, such as whether a document contains a signed consent form or whether a photo shows damage to a roof.
The second type is a multiple-choice selection. The developer offers between two and 255 named options, and the model reports a normalised probability for every candidate. The application can then pick the winner — or notice when two options are close and send the case to a human.
The third type scores against an ordered set of criteria, returning a continuous weighted value. Think of it as a rubric: how well does this support ticket match the refund policy, on a scale the business defines?
Those confidence scores let developers build a fallback into the system. Clear answers move forward automatically. Uncertain ones go to a person. Nothing is silently guessed.
Speed, Price and the Fine Print
The price is deliberately simple. Gemini Decision 1 costs $0.10 per million input tokens, and output tokens are free — because there are no output tokens. The shared input is billed only once per request, no matter how many questions are evaluated against it.
At that rate, a 5,000-token request covering a thousand records costs about $0.50 per thousand, according to VentureBeat. For bulk grading, routing and screening, the maths is hard to argue with.
Google also offers a no-cost tier in AI Studio for prototyping, with limits of 200 requests and 50,000 input tokens per minute. The endpoint runs on the Gemini Developer API’s v1beta endpoint.
There are catches. Support in Gemini on Google Cloud is listed as coming soon, not live. Question instructions are text-only, and conversation history and tool-call blocks are not supported. There is no caching or Batch API yet, so repeated context is billed in full every time.
How It Compares With OpenAI
Google is not first to this idea. OpenAI made its own Decisions API generally available on October 22, running on its GPT-6 Luna model. The race to sell fast, cheap judgement to developers is now properly on.
In Google’s own benchmarks, Gemini Decision 1 led five of the seven tests that all four compared models ran. It scored 92.2 percent on JevBench, ahead of GPT-6 Luna at 90.9 percent and Jev 1.14 at 90.0 percent.
Calibration mattered most in Google’s telling. Decision 1 recorded the lowest calibration error in the comparison, 0.0250 — roughly half of Jev 1.14’s 0.0494. When the model says it is 90 percent sure, it is right about 90 percent of the time.
For systems that act on AI answers automatically, that honesty is worth more than raw accuracy. Google also claims the API is nearly five times cheaper across the shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API.
One independent check tempers the picture. The Artificial Analysis Decision Index puts Gemini Decision 1 ninth out of 18 models at launch, scoring 65.98 at 211 milliseconds. Google’s figures are self-reported; rivals may look better on someone else’s test.
Where Google Already Uses It
The most telling detail in the announcement is that Google is already eating its own cooking. The company says its teams use the technology in three places today.
Gemini Notebook uses it to route queries to suitable models, cutting latency for everyday questions by up to 40 percent while reserving deeper reasoning for complex ones. Stitch, Google’s design tool, uses it for faster design work.
Google Search uses it as a safety layer for detecting adversarial content. That is a job where speed and calibrated confidence are not nice extras — they are the entire requirement, because a slow safety check is no safety check at all.
The launch follows Google’s expansion of access to its Gemini 4 Carbon model earlier this week. Carbon is built for complex coding, cybersecurity and financial research — long, hard tasks that need extended reasoning. The Decisions API is the opposite end of the spectrum: small, bounded judgements, made as fast and cheaply as possible.
What It Cannot Do Yet
The new API is narrow by design, and developers should not mistake it for a general-purpose model. It scores options. It does not explain itself, write summaries or hold a conversation.
That narrowness is also a limit on who can use it. There are no open weights and no self-hosting — it runs only on Google’s infrastructure. Teams that need to run models inside their own data centres will have to wait, or look elsewhere.
The missing pieces are practical. Batch processing, caching and Google Cloud support are all listed as coming, not here. Until they arrive, the API suits real-time decisions — routing a query, checking a policy — more than overnight bulk jobs.
Why This Matters for AI Agents
The bigger story is not a single API. It is the shape of the software being built on AI right now. The industry spent two years teaching models to talk. Now it is teaching them to decide — thousands of times a day, inside apps that never show a chat window.
Every AI agent that books a flight, triages a support queue or reviews a loan application makes hundreds of small judgements before it acts. Each one used to cost a full model call, a full answer and a full wait. A decision API turns those judgements into a cheap, fast, scored primitive.
That is why Google and OpenAI have shipped one within weeks of each other, and why Anthropic is pushing AI deeper into everyday work with its Claude office suite. The next wave of AI products will be judged on how reliably they choose, not how well they chat.
For developers, the pitch is simple: a thousand questions, one pass, a quarter of a second, ten cents per million tokens. If the model proves as calibrated in the wild as Google claims, the boring, repetitive judgements inside every app may soon stop being the bottleneck.
As Google’s own privacy questions around Gemini have shown, trust is earned one disclosure at a time — a lesson explored in our report on Google Gemini’s privacy changes.
The Decisions API will be judged the same way: not on its launch claims, but on what it does when a thousand real decisions are on the line. It is available now in Google AI Studio, according to The Tribune.
