
Decision Models Grab 25% of Vercel Router Traffic a Month After Launch
Decision models, invented in September, took over 25% of Vercel router requests by October. TypeSafe's Jev leads; Sqwish raises $3M; Microsoft ships Decision-1.1.
- By
- Tom Whitfield
- Filed
- Channel
- AI & Compute
- Read
- 4 min read
By Tom Whitfield
More than one in four requests on Vercel's AI model router in October went to decision models — a class of AI system invented only a month earlier. That share, measured on a real production routing service, shows how quickly a new architecture can move from research artifact to default infrastructure.
Decision models are language models that produce no output tokens. Instead, they apply general intelligence to choose quickly between preset options and register a confidence level for each choice. In production settings, a fast, cheap, well-calibrated decision often matters more than the extended reasoning of a slower frontier model.
Who leads the market?
TypeSafe built the category's anchor product. Its Jev family of models continues to dominate usage wherever public metrics exist. OpenRouter, a routing service, handled more than 1.5 billion queries to TypeSafe's models in a single week last month. That volume helped make TypeSafe a $7.5 billion company.
The technical bar to entry, however, is low. One Amazon engineer built a decision model in a weekend. Within weeks, every hyperscaler had shipped its own option. On Tuesday, Microsoft CEO Satya Nadella said the company is rolling out Microsoft-Decision-1.1 for wide use on Azure immediately.
The takeaway: the model itself is easy to copy, so the competition has moved to data and tuning.
What makes a decision model good?
TypeSafe CEO Diogo Almedia told TechCrunch that the hardest part of building Jev was finely calibrating its judgement using statistically adjusted synthetic data. The model's general intelligence came almost for free; knowing how much to trust each decision did not.
Microsoft reached the same conclusion from the hyperscaler side. Achint Srivastava, the engineering VP who led the Decision-1.1 project, said the company carefully balanced cost, latency, intelligence and calibration, and he warned against reading too much into benchmark scores.
"What we see is that 'benchmaxed' models fall flat under real enterprise usage very quickly," Srivastava wrote in an email. "Our testing underscores just how challenging it is to deliver accuracy, reliability, speed and cost efficiency in a single model. Microsoft-Decision-1.1 and Jev are among the few that perform strongly across all four dimensions, even when evaluated on 'held out' or unseen benchmarks reflective of real use cases."
Most builders in the field share that skepticism of rankings. Developers can substitute one decision model for another and test it directly in their own context, which makes live evaluation cheaper than benchmark study.
Where is the startup opportunity?
The money is also flowing to the small end of the market. Sqwish, a UK startup behind a decision model called kettei fat, announced a $3 million pre-seed round today. Cofounders Ushnish Sengupta and Federica Freddi met as Cambridge graduate students. They had been building small LLMs when Jev arrived, and they realized their architectures could be repurposed as a decision model.
Sqwish also works with synthetic data, but its differentiator is post-training tooling. Customers can upload their own production data, and the model's judgement keeps improving on the tasks that customer cares about.
"What makes us not a commodity, I think, is the fact that we are offering our customers the same power… 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," Sengupta told TechCrunch.
The kettei fat model has performed well on some benchmarks, but Sengupta frames those scores as marketing rather than proof. "The benchmark is more of an invitation to try the model out, and then the continual tuning is the reason why they should stick to us," he said.
Can a startup survive next to hyperscalers?
Freddi argues the category splits into two layers of intelligence. Future systems, she predicts, will pair System One intelligence — fast deciders like decision models — with System Two intelligence, the slower reasoning of frontier LLMs.
Even the smaller layer is a large business, she told TechCrunch: "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. This is all we care about; whereas for them, this is just a side project."
Her wager is that focus beats scale in a niche the giants treat as secondary. With Microsoft now shipping, TypeSafe valued at $7.5 billion, and continuous customer-data tuning emerging as the sticking point, the category's winners will likely be decided by calibration quality on live workloads rather than by model releases.
Source: techcrunch.com
More from Tom Whitfield
Show full bio
Staff writer covering companies, deals and strategy in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,054 articles
Related articles
ai-agents-burn-5x-the-tokens-humans-do-and-85-is-cache-rereads-d3d5f738
AI Agents Burn 5x the Tokens Humans Do, and 85% Is Cache Rereads
india-s-ai-money-chases-deployment-not-frontier-models-five-startups-to-watch
India's AI Money Chases Deployment, Not Frontier Models: Five Startups to Watch
uvi-raises-7-5m-to-sell-enterprises-their-own-context-775b28d0
Uvi Raises $7.5M to Sell Enterprises Their Own Context
openai-weighed-a-5-billion-plus-deal-to-read-the-mind-of-its-own-models
OpenAI Weighed a $5 Billion-Plus Deal to Read the Mind of Its Own Models



