Decision models are gaining ground in production AI, where a quick, cost-conscious choice can matter more than a long response. As companies from startups to Microsoft introduce models in this category, a key question is how customers can adapt them to the tasks they actually need done.
A different job for AI models
Unlike a conventional large language model, a decision model does not produce output tokens. Instead, it uses its general intelligence to select among preset options and register its confidence in that choice. That design makes it suited to situations where a fast decision is more useful than the reasoning a slower, more expensive frontier model can provide.
The category is new, but usage is already substantial. In October, more than one in four requests on Vercel’s AI model router went to decision models rather than LLMs. TypeSafe’s Jev family has been a prominent example: the company became a $7.5 billion company after releasing the models, and OpenRouter handled more than 1.5 billion queries to TypeSafe’s models last week.
Those figures suggest that developers are already trying decision models in real services. They do not, by themselves, show which model will work best for a particular task. That distinction matters as the number of available options grows.
More models, more competition
Jev’s arrival was followed by dozens of decision models. The technology can be relatively easy to build: one Amazon engineer made a model in a weekend, and hyperscalers soon had options of their own. Microsoft CEO Satya Nadella said Tuesday that Microsoft-Decision-1.1 was rolling out for wide use on Azure immediately.
Startups are entering the category, too. UK-based Sqwish announced a $3 million pre-seed round for its decision model, kettei fat. Cofounders Ushnish Sengupta and Federica Freddi had been working on small LLMs before Jev appeared. They saw that their architectures could be repurposed for decision-making.
With many models available, a general benchmark may not settle which one a business should use. Sqwish’s kettei fat has performed well on some benchmarks, but builders say developers can often substitute one model for another and see how each performs in their own setting. The practical question is whether a model makes useful, dependable choices for the work at hand.
Why customer data could matter
Training data is central to building a capable decision model. TypeSafe CEO Diogo Almedia told TechCrunch that the hardest part of building Jev was finely calibrating its judgement using statistically-adjusted synthetic data. Sqwish also works with synthetic data, while emphasizing tools that let customers use their own data to improve a model’s judgement after training.
That approach is meant to give customers a way to keep adapting a model as they collect production data. Sengupta described the idea this way: “they can keep uploading their production data, and just the same way that we trained our models, their models can also keep getting better at the tasks that they care about.”
For customers, that promise shifts attention from choosing a model once to improving its fit over time. It also gives Sqwish a way to argue that it offers more than a model that can be swapped for another. Sengupta said benchmarks may invite developers to try a model, while continual tuning gives them a reason to stay.
Performance has to work in practice
Customization is only one part of the challenge. Microsoft’s Decision 1.1 was developed by balancing cost, latency, intelligence and calibration. Achint Srivastava, the engineering VP who led the project, cautioned against judging models mainly by benchmark results.
“What we see is that ‘benchmaxed’ models fall flat under real enterprise usage very quickly,” Srivastava wrote in an email. He said delivering accuracy, reliability, speed and cost efficiency together is difficult, and that Microsoft-Decision-1.1 and Jev perform strongly across all four dimensions, including on held out or unseen benchmarks reflective of real use cases.
That leaves developers with a practical evaluation task: check how models behave in the context where they will be used, and consider whether their accuracy, reliability, speed and cost meet the need. A strong benchmark result can help prompt a trial, but the source’s examples point to real-world performance and ongoing calibration as important parts of the decision.
Sqwish sees room for smaller companies even as hyperscalers and frontier labs enter the field. Freddi expects future systems to combine System One intelligence, such as decision models, with System Two intelligence, such as frontier LLMs. For her, that division leaves a substantial opportunity for a company focused on decision models, even if the category is only one part of the work larger AI companies pursue.