Google launched the Gemini Decisions API this week, marking the third major player to ship a dedicated decision-making model that strips out reasoning tokens and charges only for input.

The API supports three question types: yes/no questions (which Google calls “Noul” questions, borrowing Jev’s terminology), multiple choice, and rubric-based scoring. If you’ve used Jev or OpenAI’s GPT-6 Luna decision model, this will feel familiar. The mechanics are nearly identical.

Pricing lands at 10 cents per million input tokens with free output, matching Luna exactly. Jev still undercuts both at 4.2 cents per million. That price gap matters if you’re running classification at scale, say, tagging support tickets or filtering content.

Why decision models exist

Standard LLM inference generates tokens sequentially, including all the reasoning steps the model works through. For tasks where you just need a binary answer or a category label, you’re paying for reasoning you don’t need to see.

Decision models short-circuit that process. They’re optimized for structured outputs where the answer itself is what matters, not the chain of thought that got you there. You ask “Is this email spam?” and get back a yes or no. No explanation, no reasoning trace, just the classification.

The tradeoff is flexibility. Jev only handles text and JSON input. Luna and Gemini Decisions support richer inputs, but all three lock you into predefined output structures. You can’t ask follow-up questions or request elaboration. These are classification engines, not conversational models.

What sets them apart

Jev launched first and remains the cheapest, but it’s text-only. That’s fine for document classification or sentiment analysis. It’s limiting if you need to process images or structured data formats.

OpenAI’s Luna and Google’s Gemini Decisions both handle multimodal inputs. Luna integrates tightly with the rest of OpenAI’s tooling. Gemini Decisions slots into Google’s Vertex AI ecosystem, which matters if you’re already running workloads there.

The practical difference comes down to latency, accuracy on your specific task, and which cloud platform you’re already using. The API contracts are similar enough that switching between them wouldn’t require a rewrite.

When this actually makes sense

Decision models shine in batch processing scenarios where you need to classify thousands or millions of items and don’t need the model to show its work. Content moderation, document routing, sentiment analysis, automated triage.

They’re less useful when the reasoning matters. If you’re building a tool that needs to explain why it flagged something, or when the decision itself is complex enough that you want to audit the logic, you’re better off with a standard model that outputs its thought process.

The other constraint is that you need to know your output schema upfront. Yes/no, pick from these five categories, score on this rubric. If your task doesn’t fit one of those patterns, or if it changes frequently, decision models add friction instead of removing it.

The broader pattern

This is the third decision model to ship in the past six months, which suggests the API pattern is sticking. The models themselves aren’t revolutionary. They’re optimized for a specific use case where most of what a language model does is overhead.

What’s interesting is that all three companies decided this use case was distinct enough to warrant its own API and pricing tier. That usually means customers asked for it enough times that building a separate product made sense.

It also means structured classification is common enough at scale that shaving costs and latency matters. If you’re processing millions of decisions a day, the difference between 4.2 cents and 10 cents per million tokens compounds quickly.

For now, Jev has the price advantage, Luna and Gemini Decisions have the feature breadth. Which one you pick depends on whether you need multimodal support and which ecosystem you’re already in.