Gemini Decision 1 picks answers instead of writing them: $0.10 per million input tokens, output is free
Google has released Gemini Decision 1, a model that doesn't write text. It answers closed questions with confidence percentages. Given an email saying "the mug arrived cracked," for example, it picks the support queue directly, so you don't pay for JSON that gets thrown away. Input costs $0.10 per million tokens, and output is free. That is a good deal for routing and checks, but you have to set the confidence threshold yourself.
AI-processed from KDnuggets; edited by Hamidun News
Picture an online store's support desk. An email arrives: "The blue mug I ordered on Monday arrived cracked on Wednesday." To send it to the right queue, most teams today call a full generative model, ask it for JSON, run the result through a parser, and pay for output tokens (the chunks of text the model splits its answer into) that are discarded immediately.
Google has decided to sell this layer separately. Gemini Decision 1 writes nothing. It reads the input and returns probabilities for closed questions such as "is the item damaged?"
and "which queue fits?" Input costs $0.10 per million tokens, and output is free because there is simply none.
If your agent has many of these quiet decision points, this is a way to make them cheaper, but with an important caveat about the confidence threshold, covered below.
What Decision 1 can do
In a single pass, the model reads text, images, video, audio and PDF, up to 262,144 input tokens per request. You can ask up to 1000 independent questions about one input, of three kinds:
- yes or no: the answer is a probability from 1% to 99%;
- a choice among 2–255 options: a probability for each;
- a rating on an ordered scale: a weighted score and a probability for each level.
Every answer comes with a confidence level, which you can use to send unclear cases to a fallback path or to a human. The shared input is billed once, no matter how many questions you ask. In exchange, the list of unsupported features is long: text generation, reasoning, function calling, structured output, batches and caching. This is an engine for scoring, not a chatbot.
Google itself already uses the Decisions API in four products. One routes requests by model tier and reasoning depth, and Google says this cuts latency on ordinary requests by up to 40%. Another builds playable interface mockups and applies voice or touch edits. The third, from Google Labs, picks an emoji reaction from 200+ options with no noticeable pause. The fourth serves as a guardrail layer and checks Model Context Protocol tools for hostile behavior.
What the KDnuggets test showed
The author ran three scenarios: the example from Google's quickstart, Google's example with an image, and an ambiguous message of his own. The model classified the cracked-mug email as damage with 99% probability and sent it to the replacements queue with 98% confidence. The image example also checked out: spilled liquid 95%, the action "wipe with a cloth" 98%, severity 1.19 out of 2. Every call produced zero output tokens.
The most telling case was the email "You charged me twice, the handle is broken." It fits two queues, and the model showed that honestly: 50.2% for replacements, 48.4% for billing, overall confidence 50%. It still rated the damage at 99%. One example is not enough to judge calibration, which is tested on thousands of cases, but the behavior you want did show up.
Here is the practical rule that follows. If your code reads only the winner, the disputed email will go to a queue as confidently as the obvious one, and nothing will be flagged. Set a confidence threshold, and send everything below it to a human or a fallback flow. Google's quickstart prints only the choice, which is not enough in production.
Where Google's claims did not hold up
Speed. Google cites 150–250 ms end to end. Text calls after the first took 339–577 ms, and network distance easily explains the gap. Requests with images took 1.1–1.4 seconds, a median five to nine times higher than claimed. Some of that time may have gone to downloading the image from a link and processing 1064 image tokens, but from the outside the causes cannot be separated. Also, Google's figures apply to contexts of up to 16 thousand tokens, not the full 256 thousand, and identical calls differed by more than 300 ms.
Comparisons. The phrase "almost five times cheaper" refers to Gemini 3.5 Flash-Lite on a benchmark set that Google chose itself. If you are already on Gemini, that is useful, but it says nothing about how Decision 1 compares with the small fine-tuned classifier you may already have running.
Limitations. The model page claims classification of live media streams, yet the Live API is on the list of unsupported features. For now, the model is available only within the Gemini ecosystem, and Google Cloud support is still marked as not ready.
What to do in practice
Decision 1 solves one narrow task well: it is cheap, answered correctly on every example, and expressed doubt when pushed to. You can try it now for request routing and rule-violation checks. Build a confidence threshold into your code from day one, and measure latency on your own images and video before you promise anyone 200 ms.
- Nov 20, 2026Google teaches Gemini to run your Mac in the background without taking over the screen or cursor
- Nov 20, 2026Google Built a Sealed-Off Copy of the Internet to Test Chrome's AI Agents
- Nov 20, 2026Anthropic shares jump 17% on first day after record $81.3 billion IPO
- Nov 20, 2026Google opens Genie 4 world model in its API: one minute of walking costs about $6
Key Questions Answered
What Decision 1 can do?
In a single pass, the model reads text, images, video, audio and PDF, up to 262,144 input tokens per request. You can ask up to 1000 independent questions about one input, of three kinds: yes or no: the answer is a probability from 1% to 99%; a choice among 2–255 options: a probability for each; a rating on an ordered scale: a weighted score and a probability for each level. Every answer comes with a confidence level, which you can use to send unclear cases to a fallback path or to a human.
What the KDnuggets test showed?
The author ran three scenarios: the example from Google's quickstart, Google's example with an image, and an ambiguous message of his own. The model classified the cracked-mug email as damage with 99% probability and sent it to the replacements queue with 98% confidence. The image example also checked out: spilled liquid 95%, the action "wipe with a cloth" 98%, severity 1.19 out of 2. Every call produced zero output tokens.
Where Google's claims did not hold up?
Speed. Google cites 150–250 ms end to end. Text calls after the first took 339–577 ms, and network distance easily explains the gap. Requests with images took 1.1–1.4 seconds, a median five to nine times higher than claimed. Some of that time may have gone to downloading the image from a link and processing 1064 image tokens, but from the outside the causes cannot be separated.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.