AI/MLNov 13, 2026

Decision models: Google joins TypeSafe, OpenAI, Microsoft race

Decision models return the best answer from a fixed list in milliseconds—and Google just entered the race.

What decision models are

The cost advantage

A rapidly forming market

Where to use them

Decision models are a fundamentally different kind of AI model. Instead of writing text token by token, they take an input and a fixed list of possible answers, then return the best one with a probability—typically in milliseconds. That makes them a natural fit for classification, routing, validation, and any task where the output must be one option from a known set. TypeSafe AI released the first decision model, Jev, on September 15, and Google's entry this week, following OpenAI and Microsoft in October, is a strong signal that this new category is more than an experiment.

The biggest practical draw is economics. Google, OpenAI, and Microsoft all charge $0.10 or less per million input tokens and nothing for output tokens, while decision models are dramatically faster than auto-regressive chatbots. The trade-off is accuracy: chatbots still score a little higher on the hardest decisions, at about 15 times the cost. For high-volume workloads where perfect accuracy is not required—or where a cheaper model can handle the easy cases first—that trade-off is often worth making.

The market is still early and unstandardized. Shanghai AI Lab reports its 4B model beat Jev on its own tests, and Microsoft built its model on Alibaba's Qwen, so there are already meaningfully different technical approaches. That means benchmarks are not yet consistent, and teams should not assume a leaderboard winner will win on their own data. The practical move is to test decision models against real inputs, measure accuracy and latency, and compare cost against the chatbot pipeline you currently use.

For developers, the most interesting signal may be adoption: on Vercel's gateway, decision models handled one in four AI requests in October. That suggests the category is already being used for ordinary production traffic, not just demos. Good first use cases are content moderation, intent classification, tool selection, response evaluation, and any guardrail that currently calls an expensive LLM to make a simple choice. The takeaway is not to replace chatbots, but to put a decision model in front of them as a fast, cheap first pass—and send only the hard cases upward.

If you're paying a chatbot to classify, route, or validate, a decision model may cut costs by an order of magnitude.

#decision models#AI infrastructure#cost optimization#TypeSafe Jev
View source ↗
Decision models: Google joins TypeSafe, OpenAI, Microsoft race | Ali Fakih