Introduction: the most interesting AI launch this week writes nothing at all
Every AI launch for two years has followed the same script. A new model arrives, it writes a little faster and reasons a little longer, and the demo shows it drafting an essay or writing a function. Then the bill arrives: you paid for a thousand words of generated text when all you needed was a yes or a no.
On November 11, 2026, Google shipped something that breaks that script. The Gemini Decisions API, powered by a new model called Gemini Decision 1, does not generate text at all. You hand it one piece of evidence — a support ticket, a screenshot, a PDF, a video clip — and up to 1,000 questions about it. It scores every candidate answer in one forward pass and returns probabilities and confidence scores — no paragraphs, no tokens dripping out, no output bill. Google quotes 150–250 milliseconds end to end, and $0.10 per million input tokens with output charged at exactly zero.
I have built enough WordPress sites with AI bolted on to know a demo from a tool. A chatbot that writes is a demo. A judge that answers a thousand checklist questions in a quarter of a second, for a fraction of a cent, is a tool.
What actually shipped on November 11
The launch is a model and an endpoint: Gemini Decision 1 (
gemini-decision-1), behind the Gemini Decisions API, available now through the Gemini API as a v1beta endpoint and in Google AI Studio, with a no-cost tier for prototyping. Vertex AI support is listed as coming soon, so today this runs on the Developer API only.It is a specialist, not a general-purpose Gemini you can chat with. A normal model decodes an answer token by token; Decision 1 scores a fixed set of candidate answers in one forward pass. You are not buying generation. You are buying judgement, measured in probabilities.
Each request carries one shared input — text, images, audio, video, or a PDF, up to 256,000 tokens — plus up to 1,000 questions about it, and the shared input is billed once no matter how many questions you attach. Google's example: a 4,000-token record with 20 checks costs about $0.50 per 1,000 records in one request, versus about $8.10 for twenty separate calls.
The three kinds of questions it answers
Everything the API does fits into three question types, and you can mix all three in one call.
First, yes/no. Is this comment spam? Does this image contain a phone number? Does this ticket need a human? You get back the probability that the answer is yes. That probability is the point. A calibrated probability is something your code can branch on: above 0.9, auto-handle it; 0.5 to 0.9, human queue; below 0.5, let it through.
Second, multiple choice: 2 to 255 options — which category does this ticket belong to, which template fits this enquiry — with normalised probabilities for every candidate. Not just the winner, but the full ranking, so you can see when it is torn between two options and route those to a person.
Third, rubric scoring: you define an ordered set of grades — a 1-to-5 quality scale, a risk scale for a submitted link — and it returns a weighted score plus the likelihood of each grade, turning a vague "is this good?" into a number with a distribution behind it.
All three run against the same shared evidence in one request. Most AI features I have shipped are not one question but a checklist — spam, safe, urgent, owner — and until now each item was a separate call, bill, and wait. Decision 1 collapses the checklist into one call.
The numbers: fast, cheap, and better calibrated than its rivals
Google's headline numbers are strong. Here is which are Google's claims and which come from independent testing.
On speed, Google reports 50–100 milliseconds of model compute for contexts up to 16,000 tokens, and 150–250 milliseconds end to end, nearly flat from one question to a hundred. The independent Artificial Analysis Decision Index v2.1 measured 211 milliseconds per task at launch, inside Google's quoted range.
On price: $0.10 per million input tokens, and output is always zero tokens, so there is no output line on the bill. No separate charges for audio transcription or OCR, and no variable thinking-token bill, because the model does not think out loud. Google claims nearly five times cheaper than the same benchmark suite through Gemini 3.5 Flash-Lite with the Interactions API. Its own cost comparison put Decision 1 at $0.88, TypeSafe Jev 1.14 at $0.68, and Flash-Lite at $4.14 — Jev is actually cheaper on that test. Decision 1 is not the cheapest classifier; it is the cheap one that also does multimodal input and a thousand questions at once.
On calibration, Google benchmarked Decision 1 against OpenAI's GPT-6 Luna, TypeSafe Jev 1.14, and Gemini 3.5 Flash-Lite, and says it won seven of nine, losing on tool-calling timing and multimodal web tasks. The number I care about is Top-1 expected calibration error: 0.0250 for Decision 1, against 0.0721 for GPT-6 Luna, 0.0494 for Jev, and 0.2104 for Flash-Lite. A 0.0250 is genuinely good — it is what makes the probabilities trustworthy enough to build thresholds on.
But here is the independent check. The Artificial Analysis Decision Index v2.1 puts Decision 1 ninth of 18 at launch, scoring 65.98 at $345 per million tasks. Ninth is not first: Google built the fast, cheap, well-calibrated workhorse, not the smartest judge in the room — and for the checklist jobs I have in mind, the workhorse is the right hire.
What Google is already using it for
Google disclosed four internal uses: Gemini Notebook routes requests to the right model tier, Stitch uses it in design work, a Labs product uses it for emoji selection, and Search uses it for safety screening.
The pattern: small, bounded judgements, made constantly, where a full generative model would be slow and expensive overkill. Model routing is the one that should make every developer sit up — I have written that logic by hand in every AI feature I have shipped, and it is always a pile of fragile rules. A 150-millisecond classifier is a better router than my if-statements.
Five things I would build with it on a WordPress site
Here is where a thousand-questions-in-one-call judge earns its keep on the WordPress and Elementor sites I build.
1. Comment and submission triage. Every site with an open form has the same problem: a queue that is 90% noise and 10% real. One call per submission answers the whole checklist — spam, abusive, sales pitch, urgency. Auto-publish the clear passes, auto-delete the clear spam, and send only the uncertain middle to a human. That middle band is the product: it is where a binary spam filter fails and a calibrated score succeeds.
2. Support ticket routing. If you run a WooCommerce store or a membership site, tickets arrive with no structure. One call can classify the ticket, score its urgency on a rubric, and detect whether it mentions a refund, a broken checkout, or an account problem. The full ranking shows when the model is torn — those are the tickets you should not auto-route.
3. Content policy checks before publishing. A scheduled job could run every new or updated post through a checklist — personal data, suspicious links, forbidden claims — in one request per post.
4. Form validation that understands meaning, not just format. A regex can tell you an email address is shaped like one. It cannot tell you the message says "please call me" while the phone field is empty, or that the enquiry is a job application sent through the sales form. A decision call can.
5. Media and accessibility checks. Because the input is multimodal, the same endpoint can check whether an image's alt text matches what is actually in the picture, or whether a product photo shows the product at all — a pre-publish safety net for Elementor builders handing sites to clients.
None of these need a chatbot or generated text — just a fast, cheap judge with honest confidence scores. That is the gap this launch fills.
The catches I would not skip
A v1beta label promises things will change, so treat this as a tool to build with, not a foundation to bet a business on. The endpoint is
POST /v1beta/decisions; the SDK call is client.decisions.create in google-genai SDK 3.2+. Pin your SDK version and wrap the call so a breaking change costs you an afternoon, not a launch.Vertex AI support is not here yet: if your compliance story depends on Google Cloud, this Developer API-only launch is not ready for you. "Coming soon" is not a date — do not plan a client deployment around it until it lands.
Context caching, the Batch API, and Flex and Priority inference are not supported — if your workload is a million documents a night, the Batch API gap matters. The free tier — 200 requests per minute, 50,000 input tokens per minute, 10,000 requests a day — is generous for prototyping, but a busy store will outgrow it by design.
Ninth of 18 means test it on your own data before trusting it with anything irreversible. Run your historical tickets and past spam through it, and measure accuracy and calibration on your cases, not Google's. A model well calibrated on average can still be confidently wrong on your weirdest 5% of inputs — and that is where your users live.
And the oldest rule in this game still applies: never let a classifier take an irreversible action on its own. Auto-delete nothing. Auto-refund nothing. Auto-ban nobody. Use the probability bands — auto-handle the certain, queue the uncertain, log everything — and let a human own the edges. The model gives you a score; what a score may do is your decision, and it should stay yours.
How to try it in ten minutes
In Google AI Studio, find the Decisions API in the no-cost tier and send one request against your own data — ten of your last support tickets, or twenty comments your spam filter already judged. Ask all three question types at once: yes/no for spam, choice for category, rubric for urgency. Then compare the probabilities against what you actually did with those items — that experiment tells you more than any benchmark, because benchmarks measure other people's data.
If it holds up, integration is small: one shared input plus questions in, probabilities out, straight into your existing thresholds. The bill is input tokens only: you pay for what you show it, never for what it says back — it says nothing back except a number.
Conclusion: the boring AI is the useful AI
The AI industry has spent two years competing on who can write the longest, cleverest answer. Google just shipped a model that refuses to answer at all — it only judges. That sounds like a smaller ambition, but most AI work on a real website was never writing. It was deciding: spam, urgent, safe, category, worth a human's time.
Decision 1 answers a thousand of those questions in one pass, in about 200 milliseconds, for $0.10 per million input tokens, with calibration good enough for thresholding. It is ninth on one independent index, still in beta, not on Vertex yet, and not the cheapest classifier you can buy. None of that changes the shape of the thing: a fast, cheap, honest judge is the missing piece between the models that write your content and the logic that decides what happens to it. I will put it in front of my own forms and queues this week, starting with the uncertain middle band where every other filter fails. If your WordPress queue never empties, try the ten-minute experiment before you buy another chatbot.
FAQ
What is Gemini Decision 1?
Gemini Decision 1 is a new Google model, launched November 11, 2026, that scores fixed candidate answers instead of generating text. You give it one shared input and up to 1,000 yes/no, multiple-choice, or rubric questions, and it returns probabilities and confidence scores for each one in a single forward pass.
How is the Gemini Decisions API different from a normal Gemini model?
A normal Gemini model generates text token by token. Gemini Decision 1 never generates text — it scores your predefined options in one pass, which makes it much faster (150–250 ms end to end) and cheaper, because output is always zero tokens and is not billed.
How much does the Gemini Decisions API cost?
On the paid tier it costs $0.10 per million input tokens, with $0 for output. The shared input is billed once per request no matter how many questions you attach. There is a no-cost tier in Google AI Studio for prototyping, with reported free-tier limits of 200 requests per minute and 10,000 requests per day.
What can I use it for on a WordPress site?
Comment and form spam triage, support ticket classification and urgency scoring, pre-publish content policy checks, meaning-aware form validation, and image or accessibility checks on uploads. Anywhere you currently run a checklist of small judgements, one call can answer the whole list.
Is Gemini Decision 1 accurate?
Google says it won seven of nine benchmarks against GPT-6 Luna, TypeSafe Jev 1.14, and Gemini 3.5 Flash-Lite, and its Top-1 calibration error of 0.0250 was the best of the group. Independently, the Artificial Analysis Decision Index v2.1 ranked it ninth of 18 at launch, so you should test it on your own data before trusting it with important decisions.
Is the Gemini Decisions API available on Google Cloud (Vertex AI)?
Not yet. At launch it runs on the Gemini Developer API (v1beta endpoint) and Google AI Studio only. Google lists Vertex AI support as coming soon, and context caching, the Batch API, and Flex/Priority inference are not supported.
Sources: Google Developers Blog, "Gemini Decisions API: Fast, calibrated multimodal decisions," November 11, 2026; VentureBeat, "Google's Gemini Decision 1 model scores up to 1,000 questions in one API call"; APIdog and Metarix launch coverage, November 2026.
