Gemini Decision 1 Tested: 54 Calls to Google's New Decisions API

TL;DR: On November 11, 2026, Google released Gemini Decision 1 (gemini-decision-1), a model that doesn't write text at all: it answers yes/no, multiple-choice and rubric questions with probabilities, through a new Decisions API. We ran 54 requests against it on a free-tier key, at no cost. It routed all 30 of our test support messages correctly, including two that tried to steer it, usually answered in about a quarter of a second, gave identical results for identical requests, and picked the right answers about a screenshot. The one trap: when an input fits none of your options, the API still picks one, sometimes with confidence near 0.8. Add an "other" option or a separate yes/no scope check, as Google's own documentation recommends.

What Gemini Decision is

Google announced the model and the API on its blog (Gemini Decisions API: Fast, calibrated multimodal decisions) and in the Google AI Studio changelog. According to the Decisions API documentation and the model page:

• You send one shared input (text, images, audio, video or PDF) and up to 1,000 questions about it in one request; the input and the questions together can take up to 262,144 tokens.
• There are three question types: yes/no (Google calls it noul; the answer is a probability between 0.01 and 0.99), choice (2 to 255 named options; you get the winner and a probability for each option) and score (an ordered rubric; you get a probability-weighted score).
• Every answer comes with a confidence value, so your code can send unclear cases to a person or a bigger model.
• The model returns zero output tokens. Instead of generating an answer word by word, it scores your options in one pass. Google's blog quotes 150–250 ms end to end.

Google lists the intended uses as routing requests to a model, queue or tool, checking content against many policy rules at once, choosing tools for an agent, and grading candidates against a rubric. The blog names internal users: request routing in Gemini Notebook, a fast canvas experiment in Stitch, emoji picking in CC from Google Labs, and a safety layer for tools in Google Search. We haven't verified any of those.

Price and limits

Rate limits are from the rate limits page on November 11; higher paid tiers get higher limits. The shared input is billed once per request, however many questions you ask; each question adds only its own instruction and option text. The documentation lists what is missing: no context caching, no Batch API, no Flex or Priority inference, and no media inside the questions themselves. Using it from the official SDK needs google-genai 3.2 or later.

How we tested

We wrote a small Python script with no dependencies that calls POST v1beta/decisions directly and saves every response. The key is a free-tier Gemini API key, so the whole run cost nothing. All 54 requests were made on the evening of November 11, 2026 (UTC), in two runs (the full set, then a second run that added the language and out-of-scope tests), one request at a time with a short pause.

The main test set is 30 short customer-support messages we wrote for this article and labelled before the first call, with one question: which team should handle it, billing, shipping, technical or other. Twenty-five are plain. Five are deliberately harder:

The last two try to steer the classifier from inside the message, a common risk when untrusted text feeds an automated system.

Results

Google's own examples: same numbers

We first ran the four examples from the documentation. All four returned the values printed there: 0.99 and 0.03 for the two yes/no questions, billing at 0.99 for the choice question, and a score of 1.12 for the rubric question. Every response reported 0 output tokens.

Support tickets: 30 out of 30

All 30 messages went to the team we had chosen, including the five harder ones. Mean confidence was 0.967. The lowest was 0.83, for the payment page that throws a JavaScript error: the model gave technical 0.83 and billing 0.14, which is a fair description of how mixed that message is. Neither of our two steering attempts worked: the "SYSTEM NOTE" message went to other at 0.96, and the "Ignore your instructions" message to technical at 0.95. Two hand-written attempts are not a robustness test, though.

The median time per request, measured by our script including the network, was 0.24 seconds. The very first call took 1.2 seconds; after that, most calls took 0.18–0.35 seconds, with a few slower ones up to 0.74 seconds. That is close to the 150–250 ms Google quotes.

Many questions in one request

We sent the same message with 1, 10 and 100 yes/no questions. Time per request was 0.24, 0.24 and 0.41 seconds; billed input was 135, 613 and 5,473 tokens. Ten questions cost no extra time; a hundred made the request about 70% slower, so in our test latency was not quite as "nearly flat from 1 to 100 questions" as Google's blog says, though still fast. A hundred checks of one message took well under half a second and about 5,500 input tokens, which at the paid price would be roughly $0.00055.

Same request, same answer

We sent one deliberately ambiguous message (a parcel that arrived with the wrong charger plug inside) five times. All five responses had exactly the same probabilities, down to the last digit: shipping 0.84, technical 0.08, other 0.08. In our run the API behaved deterministically, which makes thresholds and tests easier to rely on. Google's documentation doesn't promise this, so treat it as what we observed, not a guarantee.

A screenshot as input

We sent our own screenshot of the Antigravity CLI model picker (from our Gemini 4 Carbon article, with personal details covered) and asked four questions. It answered "is this a terminal?" with 0.94 and "is this a web browser?" with 0.05. Asked which model the menu shows as unavailable, it chose Gemini 4 Carbon with 0.9999; asked which CLI version is visible, it chose 1.5.4 with 0.999. Both are correct. Note that these were choices from short lists we supplied (six model names, four version numbers), not free reading of the text. The image counted as 527 of the request's 847 input tokens.

Other languages

We repeated the ticket task with eight messages in Ukrainian, German, Spanish, Polish and Japanese, keeping the questions in English. All eight were correct, with confidence between 0.97 and 0.998.

The trap: input that fits none of your options

This is the result worth knowing before you build anything. We sent three inputs that are not support requests at all: a trivia question ("What is the capital of Australia?"), keyboard gibberish, and a request for a poem.

With an other option, everything went to other. Without it, all three were sent to the technical team, the poem request with a confidence of 0.84, higher than our lowest correct answer in the ticket test. A rule like "accept anything above 0.8" would have let it through. With other in place, the same rule would have worked the other way: the gibberish got only 0.69 for other, so it would have gone to a person, which is what you want.

This matches the documentation: choice questions always share the probability among your options, even when none fits. Google suggests three remedies: a confidence threshold, a catch-all option with a clear description, or a separate yes/no scope question. In our small test the scope question was the clearest signal: 0.01 to 0.07 for all three off-topic inputs. Because one request can carry many questions over the same input, adding it doesn't add a second call. The documentation also warns against a key like none_of_the_above: the API sorts option keys alphabetically, so that key would come before shipping and technical instead of last.

What this means for you

• Routing and labelling. If you now send short texts to a generative model just to get a label back, we expect the Decisions API to be a cheaper and faster fit, judging by Google's pricing and our timings: no output tokens, no parsing, and a probability you can set a threshold on. We didn't run a side-by-side comparison (see the limits below).
• Always give it a way out. Add an other option or a scope check to every choice question that can receive unexpected input.
• Set thresholds on your own data. Our wrong-but-confident 0.84 shows that confidence alone is not a filter. The documentation says the same: choose thresholds with labelled examples from your own traffic.
• The free tier is enough to try it. 200 requests per minute and 10,000 per day is plenty for a prototype, but remember the free tier's data terms before sending real customer messages.
• No caching. If you ask questions about the same long document in many separate requests, you pay for it each time. Put the questions into one request instead.

Limits of our test

Our ticket set is small (30 messages plus 8 translations, five of them Ukrainian), written by us and mostly easy; it shows that the API works as documented, not how accurate it is on real traffic. The steering test had two messages and the out-of-scope test three. Everything ran on one evening, on the free tier, from a single location. We didn't compare it with a generative model on the same set, didn't test audio, video or PDF input, and didn't measure the paid tier.

How to try it

• The quickest way is the Decisions Playground in Google AI Studio: type an input, add questions, and export the code.
• The Decisions API documentation has Python, JavaScript and REST examples.
• Google also published demo apps in AI Studio, among them Doorstep Parcels and Find the Zookeeper.

What we don't know

• How Gemini Decision 1 compares with Gemini 3.5 Flash-Lite on real classification work. Google's blog says the Decisions API is nearly 5 times cheaper than Flash-Lite through the Interactions API across Google's benchmark suite; that is a cost claim, not an accuracy one, and we haven't checked it.
• Whether the identical results we saw on repeat requests are guaranteed or only typical.
• When it reaches Gemini in Google Cloud. The blog says support is "coming soon".

Sources: our own test (54 requests to gemini-decision-1 on November 11, 2026, free-tier key); Google blog, Gemini Decisions API: Fast, calibrated multimodal decisions (Nov 11, 2026); Google AI Studio: Decisions API, Gemini Decision model page, Pricing, Rate limits, Changelog.