All postsSHORT3 min read
Google's Gemini Decisions Can't Write a Word. Output Is Free
Gemini Decisions API returns odds and confidence scores with zero output tokens. Here's how it works, what input costs and where people still need to check.
Google's new AI can't write a single word. And its output is free. Google released the Gemini Decisions API on Nov 11, 2026, powered by Gemini Decision 1. You pay for input. Its answers generate zero text tokens.
What it is
The API scores the options you give it and returns odds, plus a confidence score showing how sure it is. You can ask yes/no questions, pick from 2 to 255 options, or grade against an ordered rubric.
It accepts text, images, audio, video and PDFs. You can combine them in shared input, such as a scanned claim with damage photos. It's available through the Gemini API and Google AI Studio.
How a call works
Send shared input with up to 1,000 questions in a request. The input and questions need to fit within the 262,144-token context window.
For example, send a return photo and ask whether the item arrived damaged. Add a multiple-choice question for which support queue should handle it. Those questions share the same input. The response gives you option probabilities and a separate confidence score.
Your code can route sure cases automatically and send unsure cases to a person. Google's docs show a confidence threshold of 0.7 as an example. Set your own thresholds using labeled examples from your own traffic. Keep a person on refunds and payouts in the shop setup below.
What it costs
The paid tier costs $0.10 per million input tokens. Output costs $0 because it generates zero output tokens. Shared input is billed once per request, regardless of how many questions you ask. Each question adds its own instruction and criteria tokens.
There's a free tier for prototyping in Google AI Studio. Its limits are 200 requests per minute, 50,000 tokens per minute and 10,000 requests per day. Google's pricing table marks free-tier data as “Used to improve our products: Yes.” The paid tier says “No.”
Google reports median end-to-end latency of 150–250 ms for tested multimodal contexts up to 16K tokens, with latency nearly unchanged across 1–100 questions per request. Google also says it's nearly 5x cheaper than Gemini 3.5 Flash-Lite across its shared benchmark suite. Those are Google's own benchmarks.
Start with the Google announcement, API docs and Google AI Studio. Check the pricing table and rate limits before building around the free tier.
The catch
Google's docs say user-supplied text can argue for its own classification. Treat customer input as untrusted. Write explicit criteria, test with adversarial examples and keep a person or a second check on high-stakes actions such as payouts.
It's available on the Gemini Developer API in v1beta, with Gemini Decision 1 only. It requires SDK 3.2 or later. There's no Batch API or context caching, and repeated input across requests is billed in full.
It answers closed questions. Use the regular API when you need generated text, summaries, explanations or extracted fields.
The builder angle (my opinion, no income guaranteed)
I'd sell online shops a returns-and-claims triage setup. It checks every return photo, damage claim or support ticket against the shop's own rules. Sure cases get handled automatically. Unsure cases and every refund or payout go to a person.
The deliverable I'd offer is that routing workflow, with the shop's criteria and a human review queue. Test the confidence thresholds on the shop's own labeled examples before letting code handle sure cases.
Daily email
Get each breakdown in your inbox.
One email each evening, right after the Short goes live: the AI tool it covers, the full write-up with sources, and how to get paid for it. Free, and you can unsubscribe anytime.