Gemini Decisions API: 1,000 questions for $0.10, but no free production tier in Europe

What Google has launched

Google opened the Gemini Decisions API on 11 November 2026, backed by a new model, Gemini Decision 1 (gemini-decision-1). It does not write an answer: it scores options set in advance and returns a probability for each. According to Google's announcement, it handles one shared input (text, image, audio, video or PDF, up to about 256,000 tokens) and up to 1,000 independent questions in a single pass, in 150 to 250 milliseconds end to end.

Three question types are available, according to the model page:

Google is targeting the small control decisions inside AI agents: picking a tool, checking a rule, routing a request to the right model, judging whether a context is sufficient. “Rather than generating answer text token by token, Gemini Decision 1 scores your candidate options in a single forward pass,” writes Google. Internally, Gemini Notebook already uses it to choose the model tier and thinking level for each request, serving everyday queries with up to 40% lower latency, according to the same post.

The price: input billed once, output free

Gemini Decision 1 costs $0.10 per million input tokens and nothing for output, since it produces none, according to the pricing page. The shared input is billed once per call; each question only adds the tokens of its instructions and options. Google says the model is “nearly 5x cheaper” than Gemini 3.5 Flash-Lite on its own benchmark suite, a figure no independent measurement has yet confirmed.

The saving comes from the number of calls. VentureBeat works through a 4,000-token record checked against 20 rules of 50 tokens each: in one call, that is 5,000 tokens, or about $0.50 per 1,000 records; in 20 calls that each resend the record, 81,000 tokens, or $8.10 per 1,000 records.

Two limits weigh on the bill, according to the documentation: no context caching, and no Batch API, Flex or Priority inference. The same document sent again in another call is billed in full.

In Europe: free to try, paid to serve

The point that matters for European developers is in the terms, not the announcement. A free tier exists (200 requests and 50,000 tokens per minute, 10,000 requests per day, according to the rate limits page), but the Gemini API Additional Terms of Service are explicit: “You may use only Paid Services when making API Clients available to users in the European Economic Area, Switzerland, or the United Kingdom.” Any app opened to European users must therefore run on the paid tier.

The free tier has another trade-off. The pricing page says data is used to improve Google's products on the free tier, and not on the paid tier. The terms add that human reviewers may read and annotate what is sent, and ask: “Do not submit sensitive, confidential, or personal information to the Unpaid Services.” For a prototype that sorts customer messages or PDF invoices, that rules out real data, which also keeps you clear of GDPR questions while you test.

For now, the API is only reachable through the Gemini API in Google AI Studio (v1beta). Support in Google Cloud is coming “soon”, with no date, according to Google. OpenRouter and Vercel AI Gateway also list it.

Against the other decision models

Google did not invent the category. According to Simon Willison, Gemini Decision 1 is priced like OpenAI's GPT-6 Luna decision model, at $0.10 per million input tokens, while TypeSafe's Jev remains the cheapest at 4.2 cents. Jev reads only text and JSON and OpenAI adds images; Gemini Decision 1 is the only one of the three that also takes audio, video and PDFs. The same blog notes that Gemini returns nuanced probabilities where the other two answer 1 or 0 to yes/no questions, which mainly helps if your code uses the confidence score.

For other ways AI vendors are setting terms for European users, see our guide to what Businesses in ChatGPT asks companies to sign and our explainer on OpenAI's Control Room and the EU AI Act. The EU AI Act timeline sets out which obligations apply when. More model news is in the News section.

What it changes for you

• You are building an agent: moving control decisions (tool, model, human sign-off) out of the main model cuts latency and makes the code easier to read, since it reads a number instead of parsing a sentence. Keep real access rules in your code: a probability of 0.9 is not an authorisation.
• You are prototyping in Europe: the free AI Studio tier is enough to test, as long as you keep real customer data out of it. Turn on billing before opening the service to European users.
• You process large documents repeatedly: without caching, each new call on the same document bills it again. Group your questions into a single call.

What we don't know yet

Google does not publish the detailed benchmarks behind its claim that the model leads “across the suite” (announcement), and no independent benchmark has yet measured the calibration it advertises. There is no date for Google Cloud, and we do not know whether the model will leave the beta API at the same price.