In October, more than a quarter of requests to Vercel’s AI model router went to decision-making models rather than large language models. This type of system appeared just a month earlier: instead of generating text, these models quickly and cheaply choose from a set of options and indicate how confident they are in their choice. They are used where making a decision quickly matters more than the detailed reasoning of a slower, more expensive frontier model. TypeSafe, Microsoft and UK startup Sqwish, which raised $3 million in a pre-seed round, are already competing in this market.
Companies develop models for fast decisions as startups enter the market
How decision-making models work
These models are based on large language models, but they do not generate output tokens. They are tuned to apply their general capabilities to choose one of a set of predefined options while also estimating their confidence in the decision. This approach suits operational systems where speed and cost matter, and extended reasoning is not essential.
TypeSafe became a $7.5 billion company after releasing its Jev model family. Jev models still lead in usage, at least among services that make their metrics available. Last week, for example, OpenRouter processed more than 1.5 billion requests to TypeSafe models.
Building this kind of model is relatively straightforward, so dozens of alternatives appeared after Jev launched. One Amazon engineer built a model over a weekend, and soon every major cloud provider had its own version. On Tuesday, Microsoft CEO Satya Nadella said the company was immediately beginning a broad rollout of Microsoft-Decision-1.1 on Azure.
Startup bets on customer data
UK startup Sqwish, developer of the kettei fat decision-making model, announced that it had raised $3 million in a pre-seed round. Co-founders Ushnish Sengupta and Federica Freddi met while studying for master’s degrees at the University of Cambridge. Before Jev appeared, they were developing small large language models. They later realized that their architectures could be adapted to build a model that chooses between options.
Data matters for these systems. TypeSafe CEO Diogo Almeida told TechCrunch that the hardest part of developing Jev was precisely calibrating the model’s judgments using synthetic data adjusted through statistical methods.
Sqwish also uses synthetic data, but offers built-in tools for fine-tuning on customer data. Companies can upload data from their working environments to improve the model’s ability to handle the tasks they need. Sengupta says this is what sets Sqwish apart from standard offerings: a customer’s model can keep improving on the material the company uses in its work.
Why benchmarks are not enough
Comparing decision-making models is difficult. kettei fat performed well on some benchmarks, but its developers consider tests in a specific environment more informative: replacing one model with another is usually straightforward, so its performance on a particular task can be checked directly.
Sengupta describes a benchmark as a reason to try a model, and ongoing customization as a reason to keep using it. Microsoft also developed Decision 1.1 by balancing cost, latency, intelligence and calibration. Achint Srivastava, the vice president of engineering who led the project, agreed that focusing too heavily on benchmarks is a mistake.
According to Srivastava, models optimized for benchmarks quickly perform poorly in real-world enterprise use. Microsoft’s testing, he added, shows how difficult it is to achieve accuracy, reliability, speed and cost efficiency at the same time. He described Microsoft-Decision-1.1 and Jev as two of the few models that perform well across all four measures, including on unfamiliar benchmarks designed to reflect real-world scenarios.
A market for startups
Federica Freddi believes that in future systems, type-one intelligence, which includes decision-making models, will work alongside type-two intelligence: large frontier language models. In her view, even in that future, the market for type-one systems could reach $10–50 billion. By comparison, she cited expectations around a possible OpenAI IPO and a valuation of $1–2 trillion, and noted that decision-making models could remain a side project for major players.
Dataist’s take: demand for fast decisions is already visible in Vercel’s data, but benchmarks alone cannot reliably establish the advantages of individual models. Competition will also depend on how well models perform on individual companies’ data and whether they keep improving on their tasks.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X