Gemini Decisions API: 1,000 questions a call, free output

Google is now in the decision-model race. On November 11 Google launched the Gemini Decisions API, powered by a new model, Gemini Decision 1. It does not write text. It scores your input against a fixed set of answers and returns probabilities. Input costs $0.10 per million tokens, output is free, and one call can ask up to 1,000 questions. That makes three big vendors with a decision endpoint, and that is good news for anyone paying to route, tag or check things with a chat model.

What actually happened

From Google's announcement and pricing page:

• Model: gemini-decision-1, in the Gemini API through Google AI Studio (the sample code uses v1beta). Google says Gemini in Google Cloud support is "coming soon."
• Three question types: yes/no, multiple choice (2 to 255 options) and an ordered rubric score. Every answer comes with a confidence score from 0 to 1.
• Input: one shared input of up to 256K tokens. It can be text, images, audio, video or PDFs.
• Speed: Google says 150 to 250 ms end to end for inputs up to 16K tokens, nearly flat from 1 to 100 questions per request.
• Price: $0.10 per million input tokens. The shared input is billed once per request, and each question adds only its own instruction tokens. No output, thinking, transcription or OCR charges. There is a no-cost tier in AI Studio for prototyping. The Batch API is not supported.
• Claims: Google says it leads its benchmark suite with the lowest calibration error and is nearly 5x cheaper than Gemini 3.5 Flash-Lite on that suite. The scores are in charts on Google's own page, not independent tests.

Why it matters for your business

Most of the AI work in a small business is not writing. It is deciding: is this email a refund request, which product line is this ticket about, does this listing break a policy, which model gets this job. Today a lot of that runs on chat models that bill you for every token of the answer.

Now there are three decision endpoints to choose from. OpenAI's Decisions API is $0.10 per million with 400 questions a call. Microsoft-Decision-1 is $0.042. Google is $0.10 with 1,000 questions and native audio, video and PDF input. The list price is close. The accuracy on your labels is the only number that matters, and no vendor can give you that.

What we would do this week:

1. Pull 200 real decisions your team or your current model already made, with the correct answers.
2. Run them on all three endpoints. Score accuracy, and check whether a 0.9 confidence is right nine times in ten.
3. Keep the vendor behind one function. Your question set lives in your repo, and the provider is a config value.

Sources: Google, Google AI Studio pricing.