As decision model use skyrockets, is customization is the next frontier?

Glass Epoch DeskOriginal digestSource: TechCrunch AI

AI-assisted digest, checked against the original source before publishing. How we work

Decision models drew more than a quarter of Vercel's routed AI traffic in October, and startups are now competing on letting customers tune them with their own production data.
As decision model use skyrockets, is customization is the next frontier?
Photo: Aidan Bartos · free stock license

In October, more than one in four requests on Vercel's AI model router went to decision models rather than LLMs — an enormous share of work for a kind of computerized intelligence invented only a month earlier.

Decision models are LLMs that don't produce output tokens. Instead, they are tuned to use their general intelligence to quickly and cheaply decide between preset options while registering their confidence in the choice. They suit production environments where fast decisions can matter more than the reasoning of a slower, more expensive frontier model.

The category's momentum is measurable. TypeSafe became a $7.5 billion company after releasing its Jev family of models, which still dominates usage; OpenRouter alone handled more than 1.5 billion queries to TypeSafe's models last week. Because the models are easy to build, dozens followed Jev — one Amazon engineer made one in a weekend, and every hyperscaler soon had an option. On Tuesday, Microsoft CEO Satya Nadella said Microsoft-Decision-1.1 is rolling out for wide use on Azure immediately.

The other end of the ecosystem is moving too. Sqwish Labs, the UK startup behind a decision model called kettei fat, announced a $3 million pre-seed round. Cofounders Ushnish Sengupta and Federica Freddi, who met as Cambridge grad students, had been developing small LLMs when Jev arrived and realized their architectures could be repurposed. Data appears to be the differentiator: TypeSafe CEO Diogo Almedia told TechCrunch the hardest part of building Jev was finely calibrating its judgement with statistically-adjusted synthetic data. Sqwish's pitch is built-in post-training tools that improve judgement using a customer's own data. "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 said.

Benchmark skepticism runs through both camps. Achint Srivastava, the Microsoft engineering VP who led Decision-1.1, said the project balanced cost, latency, intelligence and calibration. "What we see is that 'benchmaxed' models fall flat under real enterprise usage very quickly," Srivastava wrote in an email, adding that Microsoft's model and Jev are among the few that hold up across all four dimensions on unseen benchmarks. Sengupta made a similar point: "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."

Ranking these models remains hard, and most builders say swapping one in for another in a real context is the true test. The open questions going forward: whether customer-data tuning actually creates retention once every hyperscaler ships a free default, and whether four-way balancing of cost, latency, intelligence and calibration is defensible or just table stakes.

Key facts

  • • Decision models took over 25% of Vercel's routed AI requests in October.
  • • TypeSafe valued at $7.5 billion after Jev family launch.
  • • OpenRouter handled 1.5 billion-plus TypeSafe queries last week.
  • • Sqwish raised $3 million pre-seed for its kettei fat model.
  • • Microsoft-Decision-1.1 begins wide Azure rollout this week.

Why it matters

Builders shipping routing layers should expect decision models to absorb a growing share of cheap, latency-sensitive calls, and should treat swappability plus post-training on production data as a selection criterion rather than raw benchmark scores. As hyperscalers bundle defaults, differentiation shifts from model quality to retention loops built on continual customer-data tuning.

Spotted an error? Send a correction.

Get the day's AI news in one email

The Daily Epoch · 60-second read · unsubscribe anytime

Related