Decision Models Now Take a Quarter of Vercel's AI Routing Traffic

More than one in four requests routed through Vercel's AI model router in October 2026 went to decision models rather than large language models (LLMs). TechCrunch reported the figure.
Decision models are LLMs tuned to choose between a fixed set of preset options. They do not produce output tokens, the text fragments that generative models normally emit. Instead, they return a confidence score for the option they select. That makes them well suited to routing and classification, where the answer is one of several known labels rather than free text.
The category has a commercial leader in TypeSafe, which became a $7.5 billion company after releasing its Jev family of decision models. OpenRouter handled more than 1.5 billion queries to TypeSafe's models in the week before November 17, 2026. TypeSafe CEO Diogo Almedia said the hardest part of building Jev was calibrating its judgement using statistically adjusted synthetic data, meaning machine-generated training examples whose distribution has been corrected to match the target.
Microsoft is also moving into the segment. Microsoft CEO Satya Nadella said Microsoft-Decision-1.1 was rolling out for wide use on Azure immediately. The project was led by Microsoft engineering VP Achint Srivastava.
A smaller player made its own move the same day. UK-based AI startup Sqwish announced a $3 million pre-seed round on November 17, 2026. Its decision model is called kettei fat. Cofounders Ushnish Sengupta and Federica Freddi met as Cambridge graduate students. Sqwish offers built-in post-training tools that use a customer's own production data to improve the model's judgement.
The broader context here is where the differentiation sits. A decision model with generic calibration is a starting point, and vendors now emphasise adaptation to the customer's own traffic. Sqwish places post-training on production data inside the product, while TypeSafe's account centres on how its calibration data was built. Both are answers to the same question: how a classifier learns the judgement a specific business needs.
In my view, the customisation question matters more to platform teams than the volume figures alone. A router that sends a quarter of its requests to a confidence-scoring model is betting that routing decisions can be standardised, and that the cost of getting them wrong can be measured and corrected. Whether customisation delivers that correction in practice will show up in production metrics over the coming quarters, not in funding announcements.
For engineers, the practical takeaway is that decision models are becoming a distinct layer in the inference stack, with their own evaluation problems. Calibration, drift against production data, and handling options the model was never trained to choose between will likely shape which vendors hold their place.


