Google launched the Gemini Decisions API on Wednesday, November 11, and the model behind it does something no other Gemini model does: it writes nothing at all.

The new Gemini Decision 1 model takes one input — text, images, audio, video, or a PDF — plus up to 1,000 questions about it, and returns a probability and a confidence score for each one. No paragraphs, no summaries, no output tokens to bill. Google charges $0.10 per million input tokens, and output costs $0, because there is none.

It doesn’t write a word.

That design targets the least glamorous part of building AI agents: the thousands of small judgments that sit between a user’s request and an action. Should this ticket go to billing or support? Does this claim form have the evidence it needs? Which tool should the agent call next? A chat model answers those questions by writing sentences a developer then has to parse. Gemini Decision 1 answers with numbers a program can act on directly.

“The Decisions API gives you purpose-built primitives for bounded choices like tool selection, policy checks, model routing, and context grading,” Google DeepMind product manager Min Choi wrote in the company’s launch post.

How the Gemini Decisions API works

The API handles three kinds of questions. A yes-or-no check returns the probability that a condition holds, like whether a document contains a required disclosure. A multiple-choice question picks from two to 255 named options and reports the probability of each. A rubric question returns a weighted score across a scale the developer defines.

Every question in a request is scored in a single forward pass through the model, so the cost stays low even as the checklist grows. Google says a request can hold up to 256,000 tokens of shared input, and the model reads text, images, audio, video, and PDFs together. A claims team, for example, could send a scanned form, photos of the damage, and notes from a customer call in one request, then ask which queue should get the claim, whether evidence is missing, and how urgently someone should review it.

Speed is the other selling point. Google reports end-to-end latency of 150 to 250 milliseconds, with median model compute of 50 to 100 milliseconds, for inputs up to 16,000 tokens. Latency stays nearly flat as a request grows from one question to 100, according to the company.

The price math is where the model earns its keep. VentureBeat worked through Google’s own example: a 4,000-token record plus 20 checks costs about $0.50 per 1,000 records through the Decisions API. Twenty separate calls that each resend the same record would cost $8.10 per 1,000 records. Google says the model runs the same work for about a third of what Gemini 3.5 Flash-Lite charges per input token, and roughly 4.7 times cheaper than Flash-Lite on the benchmark workload it published.

Developers can try it now through the Gemini Developer API’s v1beta endpoint and Google AI Studio, where a no-cost tier allows 200 requests and 50,000 input tokens per minute. Support for Gemini in Google Cloud is “coming soon,” Google says, with no date attached.

Google’s benchmarks look strong, with a catch

On Google’s own tests, Gemini Decision 1 led five of the seven benchmarks that all four compared models ran. It scored 92.2% on JevBench, ahead of OpenAI’s GPT-6 Luna at 90.9% and TypeSafe’s Jev 1.14 at 90.0%. It also posted the lowest calibration error in Google’s comparison, 0.0250, roughly half of Jev’s 0.0494. Calibration matters here: it measures how closely a model’s confidence tracks how often it is right, and a confidence threshold decides how many cases get sent to a human reviewer.

Independent testing tells a more mixed story. The Artificial Analysis Decision Index v2.1, which combines six evaluations, puts Gemini ninth of 18 models at launch, with a score of 65.98 and 211 milliseconds per task. Jev 1.14 ranks higher on that index at 66.23 and answers in 92 milliseconds, while Nace AI’s Drex Max v1.1 leads the field at 72.62. Gemini did lead the index’s knowledge and reasoning tests, but on When2Call, which tests tool-use decisions, it scored 75.2%, behind Drex Max at 94.6% and Jev at 80.8%.

Google’s numbers also come with a scope note. The 150-to-250-millisecond latency figure covers inputs up to 16,000 tokens and requests of up to 100 questions. Teams that push toward the full 256,000-token window or the 1,000-question limit should measure their own files and checklists before promising that speed to users.

OpenAI got there first

Google is not first to this idea. OpenAI made its own Decisions API generally available on October 22, running on a specialized GPT-6 Luna model. It caps requests at 400 questions, against Google’s 1,000, and charges the same $0.10 per million input tokens with free output. OpenAI adds one feature Google lacks: prompt caching, at $0.01 per million cached tokens, which helps teams that send the same policy manual or product catalog across many requests.

TypeSafe’s Jev 1.14 is the cheaper hosted option for short text, at $0.042 per million input tokens, and it answers faster in independent testing. Amazon took a different route entirely: its Strands Decider 2B model is open source under Apache 2.0, so companies can run it on their own hardware and keep the decision step inside their own network. Amazon says an internal ticket-routing deployment cut average call time from about 4.2 seconds to 0.9 seconds and cost from $22,200 to $500 per million tickets after switching from Claude Sonnet 4.6.

Google does have one deployment of its own to point to. The company says Gemini Notebook already uses Gemini Decision 1 to route queries to different models and levels of reasoning effort, cutting latency by up to 40% for routine queries by reserving the heavier models for harder requests.

Why this matters

The timing is not an accident. AI agents are moving from demos into real workflows, and every workflow is full of small decisions that used to be made by people reading a screen. A model that writes a paragraph for each one is slow, expensive, and hard to audit. A model that returns a probability is fast, cheap, and easy to log.

That is also the risk. A probability looks precise even when the model is guessing, and a wrong routing decision can send a claim, a ticket, or a tool call somewhere it should not go. Google’s own guidance points the way: route ambiguous cases to a person, set confidence thresholds per decision, and run the model in parallel with the existing workflow before letting its output change a customer record.

For now, the Gemini Decisions API is a specialist tool with a clear pitch. It will not write your emails or draft your reports. It will answer a thousand small questions about a document in a quarter of a second, for a dime per million tokens, and leave the writing to someone else.

Frequently asked questions

What is the Google Gemini Decisions API?

It is a new Google API powered by the Gemini Decision 1 model that answers up to 1,000 closed-ended questions about one input in a single pass, returning probabilities and confidence scores instead of generated text.

How much does the Gemini Decisions API cost?

Google charges $0.10 per million input tokens and nothing for output tokens, since the model produces none. A free tier is available in Google AI Studio for prototyping.

How is Gemini Decision 1 different from a regular Gemini model?

Regular Gemini models generate text token by token. Gemini Decision 1 only judges: it scores every option in one forward pass, which makes it far faster and cheaper for yes/no, multiple-choice, and rubric-scoring questions.

How does Gemini Decision 1 compare to OpenAI’s Decisions API?

OpenAI’s Decisions API, generally available since October 22, runs on GPT-6 Luna, caps requests at 400 questions, and charges the same $0.10 per million input tokens. Google’s version allows up to 1,000 questions per request.

Sources: Google blog, The Decoder, Economic Times, VentureBeat