The short version: Google’s new Gemini Decisions API, released November 11, does not write anything. It answers up to 1,000 yes-or-no, multiple-choice and rating questions about a message, photo, recording or PDF in one pass, and it attaches a confidence number to every answer. That number is the part that matters to a small business: it marks the line between the calls software can make on its own and the ones that should wait for you. Input costs $0.10 per million tokens and output is free, so sorting 300 web-form leads a month costs about one cent. Three cautions in Google’s own documentation matter far more than the price, and they are below.
Most AI you have used so far is a writer. You ask, it produces paragraphs, and you read them to find the answer. Google has now shipped a model that skips the writing and goes straight to the verdict, with a built-in way of saying “I’m not sure about this one.” For an owner who spends part of every day making the same small judgment calls, that second feature is the genuinely good news here.
What is the Gemini Decisions API?
According to Google’s announcement, the Decisions API runs on a new model called Gemini Decision 1. You hand it something to look at (text, images, audio, video or a PDF) and a list of questions about it. It returns an answer to each question in roughly 150 to 250 milliseconds, a fraction of the time a chatbot takes to type its first sentence.
The developer documentation, which we read on November 12, 2026, describes three kinds of question:
- Yes or no: “Is this message a request for a quote?” The answer comes back as a probability between 0.01 and 0.99.
- Choice: “Which service is this about: lawn care, irrigation, tree work, hardscape, or none of these?” It accepts between 2 and 255 options and spreads the probability across them.
- Score: “How urgent is this, on a scale of 1 to 5?” It accepts a rubric of 2 to 255 levels, and Google recommends 2 to 10.
All the questions about one item go in a single request, and the item itself is counted once no matter how many questions you ask. Google’s own examples include checking a product photo for damage, routing a support ticket, confirming an invoice carries a purchase order number, and rating the delivery on a recorded call. Every one of those is a job that currently lands on someone at a small business, usually the owner.
How does the confidence score decide what reaches you?
Each answer carries a confidence value from 0 to 1. The documentation defines it plainly as the estimated probability that the top answer is right, and gives this instruction: “Use the answer value to pick an action and confidence to decide whether to run it automatically.”
That sentence is the whole design. Picture a landscaping company that gets 300 web-form messages a month. Most are obvious: a clear request for a mowing quote, an obvious spam pitch for search-engine services. A few are murky, like a neighbor complaining about a crew, or a commercial property manager asking about something the company may or may not do. A tool built on this model can sort the obvious ones the moment they arrive, and hold back anything below a confidence line for the owner to look at in the morning.
The software is not deciding everything. It is deciding the easy things and telling you which ones are not easy. The owner’s judgment goes where it is actually needed, instead of being spent on 250 messages that never needed it. Google’s best-practice notes say the same thing in their own terms: keep “a human or a second check on high-stakes actions such as payouts,” and route ambiguous cases to a person.
What does the Gemini Decisions API cost a small business?
Google’s pricing page, read November 12, 2026, lists Gemini Decision 1 at $0.10 per million input tokens on the paid tier, with no charge for output because the model generates none. Each question adds a few tokens of its own instructions. Using Google’s token-counting rules, here is what three ordinary workloads come to. These are our estimates, built on stated assumptions, not Google’s figures.
Sorting web-form leads. A typical inquiry of 150 words is about 200 tokens. Add three questions (is it a real job request, which service, how urgent) at roughly 40 tokens each, and a request comes to about 320 tokens. At 300 inquiries a month that is 96,000 tokens, or just under one cent.
Checking job-completion photos. Google’s token rules say an image larger than 384 pixels is cut into 768-by-768 tiles of 258 tokens each. A full-size 12-megapixel phone photo (4,032 by 3,024 pixels) is 24 tiles, or 6,192 tokens. A crew uploading 1,000 photos a month at full size would cost about 62 cents. If the tool shrinks each photo to 1,536 by 1,152 first, that is four tiles and about 10 cents a month. The documentation also lists a resolution setting without giving its token counts, so treat the full-size figure as the ceiling.
Reviewing recorded calls. Audio counts at 32 tokens per second. Four hundred three-minute calls a month is 2.3 million tokens, about 23 cents.
All three together cost less than a dollar a month. Google also claims the model is “nearly 5x cheaper” than its Gemini 3.5 Flash-Lite model on Google’s own benchmark suite; we could not check that comparison independently. For context on how Google has priced its other recent models, our look at Gemini 4 Carbon’s access and pricing covers the introductory rates that expire December 1.
The point of the arithmetic is that cost will not be the reason to use or avoid this. The confidence line will be.
Three things the documentation says that the announcement did not
1. A choice question always picks something. The documentation warns that choice questions “always distribute probability across your options, even when none fits.” Ask a five-way question about which service a message concerns, and a message about a crew blocking a driveway will still be assigned to one of the five, perhaps with apparent confidence. Google’s fixes are to add a catch-all option such as “none of these,” to ask a separate yes-or-no question first (“is this a service request at all?”), or to send low-confidence results to a person. Any tool built on this should do at least one of the three.
2. There is no correct confidence line out of the box. Google publishes no recommended threshold. Its sample code uses 0.7, 0.9 and 0.6 as illustrations only, and the guidance is to “set thresholds using labeled examples from your own traffic.” In plain terms: before trusting the sorting, someone has to run a few dozen of your real messages or photos through it, mark which answers were right, and pick the line from what that test shows. It is the same lesson we drew when a hidden checklist caught what OpenAI’s monitors missed: you measure AI against answers you already know. Note too that a yes-or-no answer never goes above 0.99, so a rule that waits for 1.0 will never fire.
3. The free tier uses your content. The free tier in Google AI Studio is generous, with 10,000 requests a day according to the rate-limits page, which notes those limits are not guaranteed. But the pricing table marks free-tier content as “used to improve our products.” Paid-tier content is not. Customer messages, job-site photos and call recordings belong on the paid tier, and at these prices there is no reason to save the cent.
Where will a small business actually meet it?
Probably not by signing up for it. The Decisions API is plumbing: it is available today only through Google’s Gemini Developer API and AI Studio, with support in Google Cloud described as “coming soon.” It does not support Google’s batch processing or caching options. A non-technical owner will most likely meet it inside software, such as an inbox tool, a field-service app or a review manager, that adds an “auto-sort” or “flag for review” feature, or through whoever builds their website forms.
That makes the useful question one for your vendor or developer rather than for Google. If a tool you rent starts sorting things for you, ask four things:
- What happens to an item the AI is not sure about? It should come to you, not be guessed.
- Is there a “none of these” path, so odd messages are not forced into the wrong bucket?
- How was the confidence line chosen, and can you see a sample of what it got right and wrong?
- Is your content running on a paid tier that does not use it to improve the provider’s products?
If you are building something yourself, the same rules apply as for any AI job. Write down exactly what each question means and what should happen with each answer, the way we described in the work order every AI job needs. A short, plain question list will do more for accuracy than any model upgrade.
Why this is good news for owners
A lot of the worry about AI in small business is about control: software doing things nobody approved. This model is built the other way around. It is fast and nearly free at the easy end, and it is designed to hand the hard end back to a person, with a number that says which is which. Used well, that means the owner stops reading 300 messages to find the 20 that need them, and starts their morning with just the 20.
Frequently Asked Questions
What is the Gemini Decisions API used for?
It answers short, structured questions about a piece of content, such as whether a message is a quote request, which service a photo shows, or how urgent a call was, and returns each answer with a confidence score instead of written text. Google’s examples include support-ticket routing, damage checks on product photos and checking invoices for purchase order numbers.
How much does the Gemini Decisions API cost?
Google lists Gemini Decision 1 at $0.10 per million input tokens on the paid tier, with no charge for output, as of November 12, 2026. By our estimate, sorting 300 web-form inquiries a month costs about one cent, and reviewing 400 three-minute recorded calls costs about 23 cents.
Is the free tier safe for customer data?
Google’s pricing table says free-tier content is used to improve its products, while paid-tier content is not. Customer messages, job photos and call recordings should go through the paid tier, which costs pennies at small-business volumes.
Can a non-technical owner use it directly?
Not easily. It is available through Google’s Gemini Developer API and AI Studio, which are developer tools, so most owners will meet it inside software they already rent or through whoever builds their web forms. The useful step is to ask that vendor what happens to items the AI is not sure about.
If software could sort one pile of small decisions for you and hand back only the unclear ones, which pile would you give it first?
