Google just shipped a model that writes nothing. Gemini Decision 1, released November 11, 2026,1 answers up to 1,000 closed questions about one input in a single request. Google's model page says it returns probabilities and bills zero output tokens. Google charges $0.10 per million input tokens for it.
One day later, Google's changelog put Gemini 4 Hydrogen on general release. Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens1, so output costs 5x input. Decision 1 removes the expensive half of the bill for any job whose answer is a label.
We have not found an independent test of Decision 1 yet. DataCamp reports that Google claims the lowest calibration error in its own comparison, ahead of TypeSafe's Jev. We could not find published scores behind that claim. What can be checked today is the price and the shape of the output.
The verdict layer under production AI
Most production AI work ends in a decision. A ticket goes to a queue. A claim gets flagged for review. Teams have usually reached those decisions by asking a chat model for a paragraph, then parsing one word out of it.
| Decision 1 input price1Google · per million tokens, zero output billed | $0.10 |
|---|---|
| Hydrogen output price1Google · per million output tokens | $7.50 |
| Questions per request1Google · closed questions on one input | 1,000 |
| Yes/no probability ceiling1Google · range runs 0.01 to 0.99 | 0.99 |
The verdict layer is the part of a pipeline that turns evidence into typed answers software can act on. Google is reportedly at least the fourth vendor selling it as a product. VentureBeat reports that TypeSafe's early-access Jev offers a cheaper way to make AI decisions, OpenAI's Decisions API runs on GPT-6 Luna and Amazon's Strands Decider lets developers host the model themselves.
Google's documentation breaks the layer into parts, and any job can be sorted into them. First comes the evidence: one shared input that can mix text and images. Together with all the questions, it can total up to 262,144 tokens1. The questions come in fixed types. A yes/no question returns a probability between 0.01 and 0.99 plus a confidence score.
A single-select question weighs 2 to 255 options and returns the top pick with a probability for each2. A rubric question rates 2 to 255 ordered levels and returns a probability-weighted expected score1. The last part is the threshold that turns a probability into an action. Google returns the numbers. Deciding what 0.82 means is your job.
How typed answers replace the parser
Picture a dock inspector with a clipboard of checkboxes. Every pallet gets the same list, and every tick comes with a note on how sure the inspector is. That's Decision 1. It reads the evidence once and ticks every box you hand it.
Google's quickstart runs it on one customer message: "The blue mug I ordered on Monday arrived cracked on Wednesday." One call asks two questions. A yes/no asks whether the item arrived damaged. A choice question asks which support queue should take it.
According to DataCamp, Google reports 150 to 250 milliseconds end to end3. That's fast enough to sit inside a live chat.
The old flow has four steps. Step one, send the message to a chat model. Step two, the model writes a paragraph. Step three, a parser hunts that paragraph for "damaged" and a queue name, and step four acts on what it finds.
Step three is where builds break. The model writes "this may possibly indicate damage" and the parser comes back empty. The verdict flow skips that step entirely. Evidence and typed questions go in. A probability and a queue name come out, ready for your code to check.
Say you triage 10,000 tickets a day at 500 tokens each. That is 5 million input tokens, or $0.50 a day on Decision 1 at Google's price. The same tickets through Hydrogen, with a 200-token answer each, cost 5 × $1.50 for input1, or $7.50. Output adds 2 million tokens at $7.50 per million, another $15.
That is $22.50 a day against $0.50, a 45x gap. And our scenario ignores Hydrogen's thinking tokens and the length of your question text. Hydrogen is a far more capable generalist. The fair question is whether this job needs its words at all.
The 1,000-question1 limit is the second lever. Send twenty policy checks as twenty separate calls and each ticket gets read twenty times, or 100 million tokens a day. Batch them into one call and each ticket is read once, for 5 million tokens. On Decision 1, that is $10 a day against $0.50.
Thresholds are where the verdict becomes a workflow. A simple rule: auto-route above 0.95, send 0.70 to 0.95 to a person, and ask the customer for a photo below 0.70. Watch the ceiling, though. Google caps yes/no probabilities at 0.99,1 so a cutoff at 0.98 leaves almost no room.
Two failures deserve a test before launch. First, one misread input skews every answer at once. If the model gets the order date wrong, all 1,000 verdicts lean the same wrong way1.
Second, calibration. A calibrated model that says 0.90 is right about 90% of the time on cases like yours. Google says Decision 1 is calibrated. Whether that holds on your tickets and your customers' languages is unclear. We think teams should treat the probability field as a claim to measure before trusting it.
Will verdicts outrun prose by 2028?
Google built this for itself first. The Decoder reports Google already uses the Decisions API for model routing in Gemini Notebook, design prototyping in Stitch and safety checks in Google Search. Routing is the telling one. Every agent step that asks a big model "which tool next?" pays generation prices for a one-word answer.
Say each agent task makes 10 routing calls with 1,000 tokens of context, and you run a million tasks a month. That is 10 billion tokens, or $1,000 a month on Decision 1. At Hydrogen's input price alone, the same reads cost $15,000 before a single output token.
Pricing points the same direction. The Decoder reports that Google puts Decision 1 at about five times cheaper than Gemini 3.5 Flash-Lite. If routing and policy checks move to verdict models, the generator becomes the last call in the chain, running only when a person needs to read words.
The risk is lopsided. A verdict on a 500-token ticket costs $0.00005, while a wrong claim denial can lose a customer and draw a regulator's attention. MIT Sloan Management Review warns that as AI output grows, people get pushed out of decisions. Cheap, fast verdicts make that drift easier to miss.
We think routing and triage move first, because those calls are reversible and cheap to audit. Hiring and credit decisions are another matter. There, a clean score can make a biased rubric look objective. No vendor has published adoption numbers, so it is unclear whether teams will move these calls by 2028 or hang on to their chat-model parsers.
Rebuild one triage step as typed verdicts
Pick one place in your stack where a chat model writes text and a parser pulls out a label. Ticket triage is the easiest place to start. The goal is one working verdict call and a calibration sheet you rerun every week.
Step one: pull 200 past tickets where you already know the right queue and whether a refund was due. Those known answers are your answer key. At 500 tokens each, the whole set is 100,000 tokens, about one cent on Decision 1.
Step two: write your questions as typed questions in Google's Decisions API. Make "Is this a refund request?" a yes/no question. Make "Which queue?" a single-select with your real queue names. Keep each instruction to one sentence, the way Google's mug example does.
Step three: run all 200 tickets and log ticket ID, question version, probability, confidence and the correct answer. A spreadsheet is enough.
Step four: check calibration. Group the yes/no answers by probability band, starting with 0.90 and up, then count how often each band was right. If the 0.90 band is right only 75% of the time, raise your thresholds.
Step five: set your bands, then try to break them. Feed it a blurry photo and a sarcastic message, and watch whether confidence drops. If confidence stays high on bad input, route that ticket type to a person.
Step six: rerun the same 200 tickets every Monday. When the 0.90 band drifts, the sheet shows it before your customers do.
Practise it step by step: Design an AI-Assisted Workflow
The bottom line
Most production AI ends in a decision, so price and test it as a decision.
Decision 1 makes typed verdicts cheap and fast, and that shifts the hard work to thresholds and calibration. The probability field is a vendor claim until your own answer key confirms it. We would start with reversible calls like triage and routing, where mistakes are easy to audit. Keep hiring, credit and claim denials under human review until the numbers on your data hold up week after week.