Not one word comes back.
The Gemini Decisions API, which Google shipped November 11, 2026, scores up to 1,000 questions in a single call and answers every one of them without generating a sentence. It runs on a new model called Gemini Decision 1 (`gemini-decision-1`). And it trades open-ended generation for closed-set decisions: yes/no checks, multiple choice, rubric scores (Google’s announcement).
One request carries a single multimodal input up to 256K tokens — text, images, audio, video, PDFs — plus up to 1,000 independent questions. Every answer arrives with a confidence score from 0 to 1. Pricing sits at $0.10 per 1M input tokens, output tokens are billed at zero. And Google pegs end-to-end latency at 150-250 ms. You can pull an API key from Google AI Studio today.
What the Gemini Decisions API Actually Returns
The model does not answer; it scores. A choice question accepts 2 to 255 candidates. That ceiling of 255 is such a specific number it reads like a buffer size somebody fought over in a design review. And returns the winner plus normalized probabilities for every option. Yes/no and rubric questions behave the same way. What comes back is a value your code can branch on, not a sentence your code has to parse and then apologize for.
The confidence score was built in on purpose.
Google’s announcement describes it as a way for your code “to route ambiguous cases to a fallback or human reviewer,” which is the feature most teams will skip and later wish they had not.
A 0.61 and a 0.97 are not the same decision, even when both pick the same answer.
My read: a big share of what gets labeled AI automation in small operations is classification dressed up as conversation. You prompt a chat model, ask it to reply YES or NO, pay for input and output, then parse a reply that sometimes shows up padded with preamble. A scoring endpoint drops the costume. Numbers survive serialization, drive thresholds, stack into dashboards. “Probably yes” does none of that.
How many prompts in your stack return exactly one word? Sit with that count for a second, because it is your migration list.
Gemini Decisions API Pricing: $0.10 per 1M Input Tokens
Cost tracks what you send, nothing else. At $0.10 per 1M input tokens and $0 for output, the bill follows the size of the request rather than the size of the reply.
Pricing pages that list $0.00 for something are my favorite genre of documentation, narrowly ahead of changelogs written in the first person.
The request shape matters more than the per-token price. Since one request carries one shared input plus up to 1,000 questions, you attach the document once and interrogate it thoroughly instead of re-attaching it for every separate question. That is where the actual savings live.
Google names four use cases for the Gemini Decisions API: tool selection, policy checks, model routing, and context grading. None of them are chat. They are plumbing decisions inside a larger system. The decisions most of us currently hand to a prompt that returns a single hopeful word.
For comparison context, a Reddit post relayed Google’s claim that the endpoint runs nearly 5x cheaper than Gemini 3.5 Flash-Lite, while correctly flagging the number as not independently verified.
Vendor benchmarks flatter vendors. Every time.
Treat the 5x as a hypothesis and measure it against your own workload before you believe it.
Gemini Decisions API Limits: 1,000 Questions, 256K Tokens, No Batch API
The unofficial details are the useful ones. That same Reddit post reports the Gemini Decisions API is v1beta only, supports text-only question instructions. And has no Batch API; none of that appears in the official sources I have. So treat it as community-reported rather than documented.
If those details hold, each one bites a different workflow. v1beta means the interface can shift under you. So put the calls behind one thin wrapper of your own instead of scattering them through the codebase.
Text-only question instructions means the input can be video or a PDF while the questions themselves must stay plain text, which rules out elaborate multimodal rubric phrasing.
No Batch API means bulk jobs run as live requests.
So count how many calls your nightly job will fire before wiring it in.
The calibration caution is bigger than all three of those.
A confidence score is only as good as its calibration, and no vendor calibrates on your data.
Before any threshold auto-approves or auto-rejects, log scores on real traffic and check whether 0.80 actually means right 80% of the time on your inputs.
That verification pass costs nothing while you prototype; a separate Reddit post reports a no-cost tier in Google AI Studio for exactly that.
Gemini Decisions API FAQ
Is the Gemini Decisions API free?
Not as such. Paid usage runs $0.10 per 1M input tokens with zero output-token charges. And a Reddit-reported no-cost tier in Google AI Studio exists for prototyping and calibration checks.
What are the Gemini Decisions API limits? Up to 1,000 questions per request on a single input of up to 256K tokens. Choice questions accept 2 to 255 candidates. Community reporting adds v1beta-only status, text-only question instructions, and no Batch API.
Can the Gemini Decisions API generate text?
No. It returns decisions and scores — yes/no values, choices with normalized probabilities, rubric results. Generation stays with the chat models.
How does the Gemini Decisions API compare to Gemini 3.5 Flash-Lite on cost? Google’s relayed claim is nearly 5x cheaper, flagged by the Reddit poster as not independently verified. Measure before you trust it.
What model powers the Gemini Decisions API?
Gemini Decision 1, served as `gemini-decision-1`, launched November 11, 2026.
What I’d Wire Up First
Start with the prompt in your stack that returns a single word. You know the one. These swaps map directly onto what Google built, and they are where the payoff is clearest:
– Triage inbound email or tickets with a bank of yes/no checks in one call, escalating anything the confidence score flags as ambiguous.
– Run policy checks across every rule at once on user-submitted content, instead of one prompt per rule per submission.
– Route requests between your own tools with a choice question, falling back to yourself when the top two probabilities sit close together.
– Grade submissions against a fixed rubric and surface only the borderline cases for human eyes.
Leave generation workflows alone.
The endpoint cannot generate, and that limit is the point.
Keep the chat model where prose belongs and hand the scorer every fork in the road. Google ships Python, JavaScript. And REST guides, plus three demo apps (Doorstep Parcels, Speed Sketch. And Find the Zoo Keeper) if you want working patterns before writing your own. The API key comes from Google AI Studio, same as the rest of the Gemini API family. And the Decisions API docs cover request anatomy.
The most useful AI endpoints are not the ones that talk.
They are the ones that slide you a number your code can act on. And the Gemini Decisions API is Google betting a whole product on exactly that idea. The move for anyone running lean automations is small and testable: pull one single-word prompt out of your stack this week, run it through Decisions on real traffic, log the confidence scores, then compare cost and accuracy against what you have now. If the calibration holds, you cut a recurring bill and sped up the decision path in one stroke.
If it does not, you learned it cheap, while nothing important was riding on the answer.
