Google launched the Gemini Decisions API on Wednesday, built on a new model called Gemini Decision 1 (gemini-decision-1). It runs through the Gemini API in AI Studio and does exactly three things: answer yes-no questions, pick from a list you supply, and rate something on a scale you define.
That is less than a language model can do, and that is the whole idea.
One forward pass instead of token by token
A normal model writes its answer one token at a time. Gemini Decision 1 scores the candidate options in a single forward pass instead, returning probabilities along with a confidence value. Everything else follows from that: 50 to 100 milliseconds of model compute, 150 to 250 milliseconds end-to-end for the caller, and latency that stays nearly flat anywhere between one and a hundred questions per request.
A single request carries one multimodal dataset of up to 256,000 tokens — text, images, audio, video or any mix — plus up to 1,000 independent questions. Pricing is 10 cents per million input tokens, and output tokens are not billed at all. Measured across Google’s own benchmark suite, the company puts the cost at just under a fifth of Gemini 3.5 Flash-Lite via the Interactions API.
Google’s examples are tool selection, policy checks, model routing and context grading. SiliconANGLE adds a concrete one: a security team asking the model to rate the severity of an antivirus alert on a scale of 1 to 10.
The category is two months old and already has three vendors
Decision models have only existed since mid-September, when the startup TypeSafe AI launched Jev, picked up a million users within days and raised $870 million at a $7.5 billion valuation three weeks later. OpenAI followed in October with a Decisions API of its own. Google is the third vendor in eight weeks, and says it benchmarked its model against both Jev and OpenAI’s service.
What this means if you build agents
Anyone building agents knows the problem. Most model calls inside a loop are not creative work. They are decisions. Which tool now, is this hit relevant, does this text break the policy, will the small model do. Paying a language model to write prose about that and then parsing the prose back out was always waste, but there was nothing else to call.
Anthropic has no comparable endpoint so far. If you work with Claude and make a lot of these micro-decisions, sooner or later you will have to do the math on whether the routing layer should stay with the same vendor as the model doing the actual work. Eight weeks ago that math did not exist.