BUSINESS
New tool · · 2 min read · Source: Google
Gemini Decisions API: 1,000 verdicts per call. You still write the rule
Google now sells judgment calls by the batch: up to 1,000 yes/no, choice or score questions in one API request, each with a confidence value. That makes lead scoring and approvals cheap, and moves the work where it belongs: to the rule.
Researched and written with AI, to Luigi's brief, positions and editorial rules. Automatically checked against the sources. How I work →
Sound familiar? Someone on the team reads 300 comments every morning, sorts leads by gut feeling, or checks ad copy against a brand PDF nobody opens anymore. None of that is creative work. It's hundreds of small verdicts. On November 11, Google shipped a tool built for exactly that.
What's new
- Gemini Decisions API, powered by the gemini-decision-1 model: up to 1,000 independent questions per request, as yes/no, as a choice among 2 to 255 options, or as a score against a rubric.
- Confidence: every answer comes with a value between 0 and 1, so your code can route uncertain cases to a human.
- Inputs: text, images, audio, video and PDFs, up to 256K tokens according to Google.
- Price and speed, per Google: $0.10 per million input tokens, output is free, 150 to 250 milliseconds latency. Available via the Gemini API and Google AI Studio with a free tier for prototyping; Google Cloud support is coming. The benchmark lead is Google's own claim.
The product isn't the AI. It's the threshold
What I find interesting: Google isn't building a smarter chat window here. It's building a switch. A comment is spam or it isn't. A lead fits the audience or it doesn't. Ad copy breaks the brand rules or it doesn't. Plus a number for how sure the model is.
Which means the real decision happens before you call the API. At what confidence does a case go through automatically? When does a human look? My rule of thumb: human in the loop belongs at the approval, not at every click. The confidence score makes that line properly adjustable for the first time.
The catch: the rubric is your brief
A score against a rubric is only as good as the rubric. If you never wrote down your brand rules, you get very fast, very cheap random verdicts. AI doesn't fix broken ways of working, it widens the cracks. An agent only knows what you give it, and that goes double for a decision model: positioning, tone of voice and no-gos need to be written as checkable criteria, not as a mood board.
One more practical point for European teams: Google's announcement names no regions and no euro pricing. Billing is in dollars, so check availability in your own market in Google AI Studio before you plan a test.
What I'd do now
- → Pick one task that today consists of many small yes/no calls: comment moderation, lead fit, claim checks on ad copy.
- → Write the rule as three to five checkable criteria. That's the real work, and it pays off even without Google.
- → Set a threshold and have a human review every case below it for two weeks. Automate after that.
Bottom line: the Decisions API makes judging cheap, so the rule you judge by becomes your most expensive asset.
Key facts
- On November 11, 2026, Google introduced the Gemini Decisions API with the gemini-decision-1 model: up to 1,000 yes/no, choice or score questions per request, each with a confidence value between 0 and 1.
- According to Google, the API costs $0.10 per million input tokens with free output, and latency is 150 to 250 milliseconds.
- It is available via the Gemini API and Google AI Studio with a free prototyping tier; Google's announcement names no regions, and Google Cloud support is listed as coming soon.
