Gemini Decisions API: test product checks before automation
No per-photo cost has been independently measured here. Validate quality and token usage on your own representative inputs before automating decisions.
Google launched the Gemini Decisions API on 11 November, giving developers a way to score predefined choices against text, images, audio, video and PDFs. For commerce teams, the useful starting point is a bounded check—such as routing a support message or flagging a product photo for review.
The Gemini Decision 1 model accepts up to 1,000 questions against a shared input, with yes/no, multiple-choice and ordered-rubric formats. It returns probabilities and confidence scores rather than writing a conventional chat response. Google lists a 256K-token input limit.
Price the actual request
Google quotes US$0.10 per million input tokens and no output-token charge. The shared input is charged once per request, while question instructions and criteria add tokens. Image resolution affects token cost, so this is not a verified per-photo price. A no-cost prototyping tier is available in AI Studio; Google Cloud support is described as coming soon.
The announcement’s 150–250 millisecond end-to-end figure is Google’s reported performance, not a service guarantee or a result from our own commerce test.
Build a review queue first
Our suggested pilot is to use a labelled set of real product images with permission to process them. Ask a narrow question, compare answers with a human reviewer, and measure false positives as well as missed problems. Keep the original image and the reason for review visible to staff.
A confidence score is not proof that a claim is true. Avoid using a photo classification alone to reject a refund, decide fraud or publish a product’s compliance claim. Start with triage and escalate uncertain cases. Compare total cost—including retries and human review—with the workflow it replaces.
This is distinct from choosing a general-purpose model for a long task. Our Stripe and OpenRouter guide covers a different part of the app-builder stack.
More from AI
See all →
