In October, more than a quarter of requests on Vercel’s AI model router were routed to decision models rather than large language models (LLMs)—a form of computer intelligence that emerged just one month ago but is already handling a significant share of work.
Decision models are also LLMs, but they do not output tokens; instead, they are fine-tuned to leverage general intelligence to quickly and cost-effectively make judgments between predefined options while simultaneously recording their confidence in each choice. They are well-suited for production environments, where rapid decision-making can be more critical than the more slow and expensive reasoning provided by cutting-edge models.
TypeSafe became a $7.5 billion company after releasing the Jev series of models, which continue to dominate usage (at least based on the metrics we can observe, such as on OpenRouter; the service processed over 1.5 billion queries targeting TypeSafe models last week).
However, decision models are actually quite easy to build; dozens of similar models emerged after Jev’s release. An Amazon engineer built one over a weekend, and soon nearly every major cloud provider had its own version. On Tuesday, Microsoft CEO Satya Nadella announced that the company is immediately rolling out its decision model, Microsoft-Decision-1.1, broadly on Azure.
This action occurs at both ends of the ecosystem: UK AI startup Sqwish today announced the completion of a $3 million pre-seed round, backed by a decision-making model called kettei fat. Co-founders Ushnish Sengupta and Federica Freddi met while pursuing graduate studies at Cambridge; they had been developing small LLMs until Jev emerged, prompting them to realize their architecture could be adapted into a powerful decision-making model.
Data appears to be key to the success of models in this category. Diogo Almedia, CEO of TypeSafe, told TechCrunch that the most challenging part of building Jev was fine-tuning its judgment capabilities using statistically adjusted synthetic data. Sqwish has also been using synthetic data, but its uniqueness lies in built-in post-training tools that leverage customers’ own data to enhance the model’s judgment.
“I think what prevents us from becoming a commodity is our ability to give customers the same capability… they can continuously upload their own production data, and just as we train our models, their models can continuously improve on the tasks they care about,” said co-founder and CEO Ushnish Sengupta to TechCrunch.
Ranking decision models is not easy; Squish’s “kettei fat” model performs well in some benchmarks, but most developers of this technology say the real test is practical performance: developers often easily swap one model for another and observe performance in specific scenarios.
“Benchmarking is more like an invitation to try out the model, but ongoing optimization is what should keep them here,” said Sengupta.
Microsoft's Decision 1.1 model is also the result of careful trade-offs among cost, latency, intelligence, and calibration. Achint Srivastava, the engineering vice president leading the project, agrees that overemphasizing benchmarks is a mistake.
“We’ve observed that models overly optimized for benchmark scores quickly become ineffective in real-world enterprise use,” Srivastava wrote in an email. “Our testing highlights how challenging it is to achieve accuracy, reliability, speed, and cost efficiency all within a single model. Microsoft-Decision-1.1 and Jev are among the few models that perform strongly across all four dimensions—even when evaluated on ‘holdout’ or unseen benchmarks that reflect real-world use cases.”
For startups like Sqwish, competing with hyperscale cloud providers and cutting-edge labs is actually an opportunity. Freddi predicts that future systems will see System One intelligence—such as decision models—working alongside System Two intelligence—such as advanced LLMs.
“Even in that future, System One might still be a $10 billion or $50 billion opportunity, while OpenAI was thinking about a $100 billion or $200 billion market cap when considering its IPO,” she told TechCrunch. “That’s all we’re focused on; for them, it’s just a side project.”




