Check, don't write: Google answers 1,000 questions in a single pass. What it means for your workflows

A woman with long copper-red curls in a black knit sweater holds her hand under a floating orb of light. In front of her are three bowls made of tracing paper; the middle one glows brightly, the two outer ones only faintly. Next to her a card: 1,000 questions in a single pass, without writing a word. Gemini Decision 1 from Google, Nov 11, 2026. Response time per Google 0.15 to 0.25 seconds, price per million input tokens $0.10, our test with Sonnet 5.5: 90 of 91 correct. An answer with a probability instead of text, Hugging Face does the same in the open

TL;DR. On November 11, Google presented Gemini Decision 1. The model answers up to 1,000 yes/no, choice and rating questions in a single pass without writing a single word. It reads text, images, audio, video and PDFs, responds in 150 to 250 milliseconds according to Google, and charges only for input. One day earlier, Hugging Face added the same class of model to its standard library, open and ready for your own server. Decision models are becoming a normal tool. Our test shows, however, that a good language model already makes such decisions almost without error. The advantage of the new class lies in speed, volume and price.

What Google and Hugging Face presented

In October we wrote about Microsoft's Decision-1 (in German). A decision model receives some content, a question and fixed answers. Instead of writing text, it gives a probability for each answer. Four weeks later, Google follows, and with a much broader scope.

According to the documentation, Gemini Decision 1 reads up to 262,144 tokens. Microsoft's model read 32,000, and only text. There are three question types: yes or no, a choice among 2 to 255 options, and a rating on a scale. Each answer comes back with a probability. The shared part of a request, such as a PDF, is billed only once, no matter how many questions are attached to it. There are no output tokens; input costs $0.10 per million tokens. Google says it already uses the model itself, for example in Search as a safety layer for agent tools.

Hugging Face followed up on November 10. Decision models are now a fixed part of Transformers, the most widely used library for open models, and they also run via llama.cpp on your own hardware. The same kind of decision is now possible without the data leaving the building.

Our test: what a writing model can already do

We did not test Gemini Decision 1 ourselves. We wanted to know what it is up against. So we wrote 24 typical requests to a machine builder, from "milling machine is down" to "dunning letter unjustified", and asked four questions about each: Which department? Is a machine down? Is it a complaint? Does the customer want a call back? The answers came from Claude Sonnet 5.5, a writing language model.

90 of 91 clear-cut decisions matched our reference. The one deviation is debatable: the model assigned a rescheduled delivery of ordered tool holders to sales instead of spare parts. At the median, each request took 1.4 seconds and 48 written tokens for four short answers.

We then also asked for its confidence, from 0 to 100, and sent each request five times. For correct answers, Sonnet stated a median of 97. For the questions we had marked as unclear ourselves, it was 70, and for the debatable deviation never more than 60.

Threshold at 80: what passes through and what a human sees

Confidence 80 or more: passes through automaticallybelow 80: goes to a human

24 requests we wrote ourselves, 4 questions each, 5 runs each, Sonnet 5.5 on November 12, 2026. A small test against our own reference, not proof for your data.

So the threshold already works with a writing model, at least in this small test. But it costs extra: with the confidence, Sonnet wrote 115 instead of 48 tokens, and each request became a good third more expensive. A decision model delivers the probability at no extra charge. Google also promises that it is calibrated, meaning that "90 percent" really means 9 out of 10 hits.

What it costs and how fast it is

Using the token counts from our test, we calculated what 10,000 such emails would cost at list prices.

Classifying 10,000 emails with 4 questions each, in dollars

Our own calculation with the token counts from our test (median 425 input tokens and 48 output tokens per email) and the list prices from Anthropic and Google as of November 12, 2026. Each model counts tokens slightly differently. Bars drawn to the same scale.

Compared with a large language model, the gap is wide, about 31 times. Compared with Google's small Flash-Lite, it is almost six times, which fits Google's claim of "nearly 5x cheaper". Against Anthropic's small Haiku 5.5, only a factor of 1.6 remains. If you already use a small model today, switching will save you very little money.

The difference in speed is clearer. A writing model needs more time for every additional question because it writes out every answer. According to Google, Decision 1 stays almost equally fast, whether it gets one question or a hundred.

Response time per request, in seconds

Sonnet measured through our own model gateway, including the network path. For Decision 1, Google states 150 to 250 milliseconds for the whole request. These are not the same conditions.

The catch

  • German is not documented. Google does not list any languages. Whether the probabilities are as reliable in German as in English is something everyone has to check for themselves.
  • Data and location. So far, Decision 1 only runs through Google's developer API; Google Cloud is to follow "soon". The documentation names no data location. On the free tier, Google uses the content to improve its products; on the paid tier, it does not.
  • Google's numbers are Google's numbers. Google runs the comparison on its own test suite. The text contains neither hit rates nor calibration figures.
  • No reasoning, no context. The model says what it chooses, not why. The questions in a request cannot see each other. For a choice question, it always spreads the probability across the options offered, even if none fits. That is why an "other" option belongs in every list.
  • The open models are young. Hugging Face gives no measurements and itself calls the field fast-moving.

What it means for mid-sized companies

What is new above all is what can now be checked. With images, PDFs and audio, the technology no longer fits only the inbox. A supplier invoice as a PDF can be checked against a list of mandatory details. A photo from goods receipt can be asked: Is the label legible? Is the packaging damaged? A recording of a service call: Was an appointment made? Is a call back needed? Many companies ask such questions hundreds of times a day, and today a person or an expensive model answers them.

For small volumes, little changes. If you handle a few hundred cases a day, a good language model plus a request for its confidence will continue to serve you well. If you work with sensitive data, the open models let you run the technology in-house.

What you should do now

  1. Count your decisions. Where in your workflows does someone only sort, check or tick boxes, and how often a day? A dedicated rebuild only pays off at large volumes.
  2. Ask closed questions. Yes or no, a choice, a scale. Always with an "other" or "unclear" option.
  3. Measure with your own cases, in your language. 100 to 200 cases that your people have assigned. Count the hits and check whether high confidence really means correct in your case. Our test with 24 cases took less than an hour.
  4. Set the threshold. Above it, the case moves on; below it, a person takes a look. In our test, it was 80.
  5. Settle the data route. Google's paid tier, later Google Cloud, or an open model on your own hardware. Your data protection decides this, not the price.

In most workflows, decisions are made far more often than text is written. For that, there is now a dedicated tool from Microsoft, from Google and as an open model. If you want to know which decisions in your company are suited to it, we are happy to work it out with you on your own cases.

A conversation worth having, even without an engagement.

20 minutes, no obligation. Usually a personal reply the same working day.