Google releases Gemini Decision 1: 1000 questions at once for $0.10 per million tokens

Here is a familiar situation: you need to determine whether a photo shows a stain, whether a user's request is safe or which model should handle it. You send it to a large LLM, it produces a paragraph of reasoning, and you then extract the word “yes” with a regular expression. It is expensive, slow, and gives you little sense of how confident the model is in its answer. Google has addressed this problem with a dedicated product.

What Google released

Google DeepMind has introduced the Gemini Decisions API. It is an interface for three types of closed-ended questions: yes/no, multiple choice and ratings on a scale. Under the hood is the Gemini Decision 1 model. Instead of generating text, it directly scores the options and returns probabilities along with a confidence value. Responses take 150–250 milliseconds.

Its main feature is the ability to process up to 1000 independent questions in a single pass. Inputs can include text, images, audio, video and PDFs. In other words, a single document or frame can be run through an entire questionnaire at once.

What it looks like in practice

Google demonstrates it with a photo of a kitchen countertop. In a single pass, the model answers questions of all three types:

• Is there a stain? “Yes,” with 95% confidence;
• Which cleaning action is appropriate? An option selected with 98% confidence;
• How dirty is it? 1.19 on a scale from 0 to 2.

In a conventional setup, this would require three separate LLM requests and parsing the text responses. Here, a single call returns numbers that can feed directly into application logic: if confidence falls below a threshold, send the frame to a human.

Google already uses the API internally. It handles dynamic model routing in Gemini Notebook, helps with design prototypes in Stitch and performs safety checks in Google Search. Routing and moderation call for precisely this kind of fast, inexpensive decision about where to send something or whether to let it through, rather than a lengthy answer.

Pricing and benchmarks

Pricing is $0.10 per million input tokens, with no charge for output tokens. Google says this is roughly five times cheaper than Gemini 3.5 Flash-Lite. Access is already available through Google AI Studio, while integrations with OpenRouter, LangChain, Vercel and other platforms are rolling out.

According to Google, Gemini Decision 1 leads a suite of benchmarks covering agent routing, reasoning and multimodal understanding. It also has the lowest calibration error (Top-1 ECE). Put simply, when the model says it is “90% confident,” it is actually correct about 90% of the time. For an API that returns probabilities, this matters more than an extra percentage point of accuracy: decision thresholds work only when the numbers are trustworthy.

Similar models have been appearing one after another in recent weeks. Startup Jev introduced its version first, followed quickly by OpenAI, Cloudflare and Microsoft. “Decision models” appear to be emerging as a distinct class of tools alongside chat models.

Limitations and next steps

All the figures currently come from Google, including the claim of being “five times cheaper” and the benchmark leadership. The article includes no independent verification. Based on the description, the model is suitable only for closed-ended questions. If you need an explanation or a detailed answer, you will need a conventional LLM. Some integrations are still pending.

Practical steps for teams:

• Identify places in your product where an LLM serves as a classifier: request routing, moderation, data labeling and attachment checks.
• Build a small labeled dataset from your own data and compare Decision 1 with your current setup on cost, latency and accuracy.
• Test calibration separately on your own data. If the probabilities are reliable, you can build cascades: the model handles high-confidence cases, while uncertain ones move to the next stage.

If Google's claims hold up, inexpensive decision models could take over all routine classification from large LLMs, leaving the larger models to handle tasks that require thinking and writing.