AI & Models
Decision models go mainstream, from Azure to a two-person startup
More than one in four AI requests on Vercel's router were decision models rather than full-size language models in October.
Decision models are large language models tuned to skip generating output text and instead pick quickly and cheaply among a set of preset options, while also reporting how confident they are in the choice — useful in production systems where speed matters more than the depth of reasoning a frontier model would otherwise provide. TypeSafe, now valued at $7.5 billion after releasing its Jev family of decision models, continues to dominate the category: by OpenRouter’s own count, TypeSafe’s models fielded more than 1.5 billion queries just last week.
Because a decision model is comparatively simple to build, the category has multiplied fast since Jev’s debut: dozens were released in its wake, one Amazon engineer built one in a weekend, and soon every hyperscaler had its own. On Tuesday, Microsoft CEO Satya Nadella said the company is rolling out its version, Microsoft-Decision-1.1, for broad use on Azure immediately. The same day, UK-based startup Sqwish announced a $3 million pre-seed round for its decision model, “kettei fat.” Co-founders Ushnish Sengupta and Federica Freddi, who met as Cambridge graduate students, say they were building small language models when Jev arrived and realized their architecture could be repurposed for the same job.
Builders in the space describe data, not raw benchmark scores, as the real differentiator. TypeSafe CEO Diogo Almeida said the hardest part of building Jev was calibrating its judgment with carefully adjusted synthetic data; Sqwish works with synthetic data too but is leaning on tools that let customers keep feeding in their own production data to keep improving the model after deployment. Achint Srivastava, the Microsoft engineering VP who led the Decision-1.1 project, likewise argued that models tuned too tightly to published benchmarks tend to underperform in real enterprise use, and said Decision-1.1 and Jev stand out for holding up across cost, latency, intelligence and calibration even on benchmarks the companies didn’t train against.
Freddi frames the moment as an early split between fast “System One” decision models and slower, more capable “System Two” language models working in tandem. “Even in that future, System One is maybe a $10 billion, $50 billion [opportunity], and OpenAI is doing this IPO with you know $1 trillion, $2 trillion market value in mind,” she said. “This is all we care about; whereas for them, this is just a side project.”
Why it matters
A category invented barely a month before this reporting is already routing more than a quarter of one major platform’s traffic, with a hyperscaler and a two-person startup moving on the same day — a sign cheap, fast decision-making may become as standard a layer of AI infrastructure as the frontier models that get the attention.