A lot of AI work inside a business is not writing. It is small decisions made thousands of times a day: which team should get this ticket, is this lead worth a call, does this message try to trick our chatbot, is this review positive. Using a full chat model for each of those calls is slow and expensive, a bit like hiring a writer to tick boxes.
On November 11, 2026, Google launched a model built only for those boxes. Gemini Decision 1, available through the new Gemini Decisions API, does not generate any text. It reads an input, answers up to 1,000 typed questions about it in one pass, and returns probabilities and a confidence score. It costs $0.10 per million input tokens, and output is free because there is none. That puts Google directly into the decision model market that TypeSafe AI opened with Jev in September and OpenAI joined in October.
This guide explains what Gemini Decision 1 is, what it costs and where its limits are, what Google's benchmarks show, how it compares with Jev and OpenAI, and what happened when we ran our own test on 26 marketing and e-commerce questions in English and Vietnamese.
What Is Gemini Decision 1?
Gemini Decision 1 is a decision model. Instead of writing an answer word by word like Gemini, ChatGPT or Claude, it “scores your candidate options in a single forward pass, returning probabilities and confidence scores,” according to Google. The model ID is gemini-decision-1.
It answers three types of questions, which Google's Decisions API documentation names as follows:
- noul (yes or no): returns the probability that a condition is true, always between 0.01 and 0.99. Example: does this message try to override the assistant's instructions?
- choice: picks one option from 2 to 255 choices and returns a probability for every option. Example: which support team should handle this ticket?
- score: places something on an ordered scale of 2 to 255 levels (Google recommends 2 to 10) and returns a weighted score. Example: how positive is this review, from very negative to very positive?
Every answer also comes with a confidence value from 0 to 1, Google's estimate of the probability that the top answer is right. That lets software act automatically on clear cases and send unclear ones to a person.
Two things set it apart from earlier decision models. First, it is multimodal: one request can include text, images, audio, video and PDFs, so a single call can check a scanned invoice, a product photo and a customer note together. Second, one request can ask up to 1,000 questions about the same input, and the shared input is billed only once.
How Is It Different From a Normal Gemini Model?
A normal large language model is built to write. Gemini Decision 1 cannot. Google's model page lists text generation, function calling, structured outputs, thinking, search grounding and image generation as not supported. What it gives up in flexibility, it gains in speed, price and predictable output:
- Zero output tokens. You pay only for what you send in, so there are no surprise bills from long answers or hidden reasoning.
- Fixed answer format. The result is always a probability, a choice or a score, so there is no text to parse and no risk of a malformed reply.
- Speed. Google says median model compute is 50 to 100 milliseconds, or 150 to 250 milliseconds end to end, for inputs up to 16,000 tokens, and stays nearly flat from 1 to 100 questions per request.
- Calibrated probabilities. Google says the probabilities are calibrated, meaning a 0.8 answer should be right about 80% of the time, which is what makes confidence thresholds useful.
How Much Does Gemini Decision 1 Cost, and What Are the Limits?
- Price: $0.10 per million input tokens across text, images, video, audio and PDF, and $0 for output. There is a free tier in Google AI Studio for prototyping.
- Context window: 262,144 tokens, shared across the input and all questions in a request.
- Rate limits: the free tier allows 200 requests per minute and 10,000 per day. Paid tiers start at 1,500 requests per minute and go up to 15,000, according to Google's rate limits page.
- Not supported: context caching, the Batch API, and Flex or Priority inference. Google notes that repeated input across requests is billed in full each time.
- Availability: the Gemini Developer API (v1beta) and Google AI Studio, including a Decisions Playground for testing without code. Google says support in Gemini in Google Cloud is “coming soon”. It is also available through OpenRouter, which is how we tested it.
For comparison, Google's Gemini 3.5 Flash-Lite costs $0.30 per million input tokens and $2.50 per million output tokens. Google says Decision 1 is “nearly 5x cheaper” than Flash-Lite across its benchmark suite, mostly because there are no output tokens to pay for.
What Do Google's Benchmarks Show?
Google's benchmark table is published as an image. The scores below come from DataCamp and match those reported by 36Kr. The order is Gemini Decision 1, TypeSafe's Jev 1.14, OpenAI's GPT-6 Luna on its Decisions API, and Gemini 3.5 Flash-Lite. All are Google's own numbers and have not been independently reproduced.
- JevBench, TypeSafe's own decision benchmark: 92.2% vs 90.0% vs 90.9% vs 86.6%.
- Amazon ESCI, product search relevance (Macro-F1): 58.7 vs 56.7 vs 54.7 vs 47.2.
- New Yorker Caption, matching captions to cartoons: 79.7% vs 72.0% vs 68.4% vs 73.7%.
- PubMedQA, biomedical questions: 79.8% vs 78.2% vs 74.4% vs 77.6%.
- Global-MMLU-Lite, multilingual knowledge: 86.0% vs 85.0% vs 79.6% vs 78.8%.
- Mind2Web Multimodal, choosing the next action on a web page: 86.2% vs 86.8% vs 81.8% vs 78.0%. Jev wins.
- When2Call, deciding whether and how to call a tool: 79.9% vs 82.0% vs 70.4% vs 60.9%. Jev wins.
- MMAU, audio understanding: 71.8% vs 67.7% for Flash-Lite. Jev and GPT-6 Luna have no score.
- Video-MME, video understanding: 69.1% vs 63.3% for Flash-Lite. Jev and GPT-6 Luna have no score.
Google also reports the lowest calibration error of the four: 0.025, against 0.049 for Jev, 0.072 for GPT-6 Luna and 0.210 for Flash-Lite. On cost, DataCamp reports the total for the seven shared tests as $0.88 for Gemini Decision 1, $0.68 for Jev, $0.93 for GPT-6 Luna and $4.14 for Flash-Lite.
Read the table with three caveats. Two of Google's seven wins, audio and video, have no competitor scores because Jev and GPT-6 Luna cannot take those inputs. The two losses are both to Jev, and both are on agent tasks such as tool selection, one of the main jobs Google is pitching the model for. And Jev is still the cheapest of the three decision APIs on Google's own cost figures.
Gemini Decision 1 vs Jev vs OpenAI Decisions API
- Gemini Decision 1: $0.10 per million input tokens, free output, 262,144 token context, text, images, audio, video and PDF, up to 1,000 questions per request.
- TypeSafe Jev 1.14: $0.042 per million input tokens, free output, a default context of 64,000 tokens per request (262,000 with an opt-in setting), and text plus experimental image input, according to TypeSafe's documentation. It does not accept audio or video. Jev launched in mid-September and TypeSafe later raised $870 million at a $7.5 billion valuation, according to SiliconANGLE.
- OpenAI Decisions API with GPT-6 Luna: $0.10 per million input tokens and free output, reading text and images, according to eesel AI.
In short: Jev is the cheapest per token and stronger on some agent tasks. Gemini Decision 1 is the only one that reads audio, video and PDFs natively, and it leads most of Google's accuracy and calibration tests. OpenAI's offer sits in between on price but trails on Google's numbers.
Our Test: 26 Marketing and E-commerce Decisions
Vendor benchmarks only go so far, so we ran our own small test on November 13, 2026, calling both Gemini Decision 1 and Jev through OpenRouter's decisions endpoint from Vietnam. We wrote 26 questions with known correct answers, drawn from the kind of work our clients do, with 5 of the 26 written in Vietnamese:
- Ticket routing (choice, 8 questions): send each customer message to billing, technical, shipping or returns.
- Lead qualification (yes or no, 6 questions): is this a real lead for a Shopify and e-commerce marketing agency?
- Prompt injection (yes or no, 6 questions): does this message try to override the assistant's instructions?
- Review sentiment (score, 6 questions): rate each review from very negative to very positive, including one sarcastic review and one mixed review.
The results:
- Accuracy: both models got 25 of 26 right, including every Vietnamese question and the sarcastic review. Both missed the same one: “Good quality, but shipping took longer than promised,” which we labeled positive and both rated closer to neutral.
- Confidence where it counts: when Gemini Decision 1 was wrong, its confidence was 0.50, a clear signal to send the case to a person. Jev was wrong with 0.72 confidence. On correct answers, both averaged about 0.92 to 0.93.
- Speed: median response time was about 820 milliseconds for Gemini Decision 1 and 700 milliseconds for Jev, measured end to end from Vietnam through OpenRouter. That is well above Google's 150 to 250 millisecond figure, which likely reflects a direct connection closer to Google's servers, so test from your own location.
- Cost: all 26 calls cost $0.00034 on Gemini Decision 1 and $0.00037 on Jev. Jev is cheaper per token, but it counted 8,699 input tokens for the same content where Gemini counted 3,441, so the cost per call ended up almost identical, about $0.000013 to $0.000014.
This was a small test of fairly clear cases, run once, and it used text only, so it says nothing about Gemini Decision 1's image, audio or video skills. Its main lessons are practical: both models handle routine marketing decisions well, including in Vietnamese; price per token is misleading when models count tokens differently; and Gemini's confidence score was the more useful warning sign on the one case both got wrong.
What Can Marketing and E-commerce Teams Use It For?
- Support triage: route emails and chats to the right team and flag urgent or angry messages.
- Lead scoring: qualify form fills and inbound emails before they reach sales.
- Review and comment tagging: score sentiment and tag topics such as shipping, sizing or price across thousands of reviews.
- Chatbot guardrails: check every incoming message for prompt injection, and every outgoing reply against your policies, before it is shown.
- Product and ad checks: because it reads images and PDFs, it can check whether a product photo meets your guidelines or whether an ad creative includes required wording.
- Model routing: send easy questions to a cheap model and hard ones to a strong model. Google says its Gemini Notebook product uses Decision 1 this way, with up to 40% lower latency on everyday queries.
If you are building these flows with AI agents, decision models are the fast, cheap step that sits between a trigger and a full language model, a pattern we cover in our guide to AI orchestration.
What Are the Limits and Risks?
- It only answers questions you define. Anything open ended, such as writing a reply, still needs a language model.
- A choice question always spreads probability across your options, even when none of them fits. Google recommends adding a catch-all option such as “other”.
- The right format does not mean the right answer. Google's own advice is to set confidence thresholds from your own labeled data and keep a human check on high stakes actions.
- Treat inputs as untrusted. Customer messages can contain instructions aimed at the model.
- Results vary by type of input. In hands-on testing by MindStudio, the model did well on text, images and PDFs but misread the emotion in an audio clip and struggled to spot an AI-generated video.
- It is a v1beta API without caching or batch discounts, and it is not yet available on Google Cloud, which matters for companies that need enterprise contracts.
How to Get Started
- Try it without code in the Decisions Playground in Google AI Studio.
- Pick one high volume decision your team makes by hand, such as ticket routing, and collect 50 to 100 real examples with the correct answers.
- Write each question in plain language and give every option a short description. Keep score scales simple, such as 3 to 5 levels.
- Run your examples, check accuracy, and choose a confidence level below which cases go to a person.
- Compare cost per call, not price per token, against Jev or a small language model before you commit.
Decision models are one of the most practical AI tools of 2026 for businesses, because they automate the many small calls that slow teams down without the cost of a full chatbot. If you want help choosing the right decisions to automate and setting them up safely, our AI training program can help. Contact us to plan a session.
Frequently asked questions
What is Gemini Decision 1?
Gemini Decision 1 is a Google decision model launched on November 11, 2026. Instead of generating text, it answers yes or no, choice and score questions about an input of text, images, audio, video or PDFs in a single pass, and returns probabilities with a confidence score.
How much does Gemini Decision 1 cost?
It costs $0.10 per million input tokens and nothing for output, because it produces no output tokens. There is a free tier in Google AI Studio for prototyping, limited to 200 requests per minute and 10,000 per day.
Is Gemini Decision 1 better than Jev?
On Google's own benchmarks, Gemini Decision 1 beats Jev on five of the seven tests both took and has lower calibration error, while Jev wins on two agent tasks and is cheaper per token. In our 26 question test, both scored 25 of 26 and cost about the same per call. Gemini also reads audio, video and PDFs, which Jev does not. See our full [Gemini Decision 1 vs Jev comparison](/blog/gemini-decision-1-vs-jev).
How fast is Gemini Decision 1?
Google says 150 to 250 milliseconds end to end for inputs up to 16,000 tokens. In our test from Vietnam through OpenRouter, the median was about 820 milliseconds, so real speed depends on your location and connection.
Can Gemini Decision 1 write text or call tools?
No. It does not support text generation, function calling, structured outputs or thinking. It only returns answers to the yes or no, choice and score questions you define.
Does Gemini Decision 1 work in Vietnamese?
In our test, it answered all 5 Vietnamese questions correctly, covering ticket routing, lead qualification, prompt injection and review sentiment. Google also reports 86.0% on the Global-MMLU-Lite multilingual benchmark.
Need help with this?
Growthtrait can help you put this into practice. Let's talk about your goals.
Contact us




