
Google has a new toy for developers, and on paper it is genuinely useful. The Gemini Decisions API does not write essays or chat. It makes up its mind. Point it at some text, an image, a video or a PDF and it will answer up to a thousand yes/no, multiple-choice or rubric questions about it in a single pass, each answer stamped with a confidence score, for about ten cents per million input tokens. That is clever engineering. It is also Google doing the thing Google always does: positioning itself as the layer every other piece of software has to run its decisions through.
What it actually does
It does closed-set classification at scale. Instead of asking a chatbot whether a review is positive and paying for a paragraph of waffle in reply, you hand the new Gemini Decision 1 model a thousand such questions at once and get back scores rather than prose. You are billed for the input and nothing for the output. It handles text, images, audio, video and PDFs, and Google claims answers in a couple of hundred milliseconds, which is a vendor figure so treat it as marketing until someone independent checks. It launched on 11 November, and OpenAI made its own Decisions API generally available back in October, so this is already a two-horse race.
The honest use case
Credit where it is due: this is a legitimately handy building block. If you run any pipeline that sorts things, moderating content, triaging support tickets, tagging documents, grading answers, pulling structured yes/no facts out of a stack of PDFs, this is cheap, fast and fit for purpose. For high-volume, low-drama classification it is a good deal today, and the free tier in Google AI Studio means you can test it for nothing.
Now the catch, and there is always a catch
Three of them. First, at launch there is no Google Cloud support, no context caching and no batch processing, so the thousand-questions-at-once headline comes with asterisks the moment you try to run it in production. Second, the confidence score invites misplaced trust. A number between zero and one looks like certainty, but it is a model’s guess about its own guess. Automate a thousand decisions a second and you also automate a thousand mistakes a second, except now each error arrives wearing an official-looking score. Third, and this is the one that matters, you are wiring your product’s judgement directly into Google’s model.
Remember who you are handing the decisions to
This is a company that a US court found to have run an illegal monopoly in search, that is under antitrust pressure across several continents, and that built its empire by making itself the unavoidable middleman and then charging rent for the privilege. A product literally called Decisions is that same playbook aimed at the next layer up: not only answering your searches, but making the small automated judgements inside everyone else’s software. Ten cents a million is an introductory price. The toll tends to rise once you have built your whole pipeline on it and unpicking it would cost you a quarter of engineering time. Cheap, fast and sticky is the oldest trick in the Google catalogue, and it has worked every time.
Should you use it
Yes, with your eyes open. Keep an abstraction layer so you can swap in OpenAI’s version or an open model without rewriting everything. Never let a confidence score stand in for a real decision on anything that actually matters, money, safety, people. And assume today’s price is a loss-leader designed to get you dependent. The tool is good. The landlord is the problem.
What this means
The Decisions API is a useful, well-built piece of infrastructure and a small land-grab at the same time. Google has worked out that the valuable position is not writing the answer, it is being the thing that decides. Use the tool for what it is good at, pay the ten cents, and keep one hand on the exit, because the whole history of this company says the exit gets narrower the longer you stay.
Edgar Friendly, calling the AI bullsh*t. More tools, more swearing: toptoolstack.com.
Did you know? Google and OpenAI both now sell a “Decisions” API, launched weeks apart. When the two biggest players race to own the same new layer at once, it is usually because they have both spotted the same toll booth.
Sources: Implicator.ai; OODAloop; Interesting Engineering; ExplainX.