A new kind of AI model barely existed two months ago. In October it handled more than one in four requests on Vercel’s AI model router. TechCrunch AI reports that decision models are now spreading fast across production AI, and the next fight is over who lets customers tune them.
Decision models are LLMs that don’t write text. Instead, they use their general intelligence to pick between preset options quickly and cheaply, and they report how confident they are in each pick. When you’re routing a support ticket, flagging a transaction or choosing which tool an agent calls next, you need a fast, cheap answer more than a long explanation.
📈 From Zero to Default in One Month
The category moved very fast. According to TechCrunch AI, TypeSafe became a $7.5 billion company after it launched its Jev family of models. Jev still leads usage: OpenRouter alone handled more than 1.5 billion queries to TypeSafe’s models last week.
The moat looks shallow, though. Dozens of competitors followed Jev, and one Amazon engineer reportedly built a decision model in a weekend. Every hyperscaler now has one. On Tuesday, Satya Nadella said Microsoft-Decision-1.1 is rolling out for wide use on Azure immediately.
At the other end of the market, U.K. startup Sqwish raised a $3 million pre-seed round for its model, kettei fat. Co-founders Ushnish Sengupta and Federica Freddi met as Cambridge grad students and were building small LLMs when Jev came out. They realized their architectures could be repurposed for this job.
🎯 Calibration Is the Product
What stands out here is where builders say the hard work lives. It’s not the architecture. It’s the data. TypeSafe CEO Diogo Almeida told TechCrunch the toughest part of building Jev was calibrating its judgment with statistically adjusted synthetic data.
Sqwish is betting on that point. Its pitch is built-in post-training tools that keep improving the model with a customer’s own production data. “What makes us not a commodity, I think, is the fact that we are offering our customers the same power,” Sengupta said. In his words, the benchmark is “an invitation to try the model out,” and continual tuning is “the reason why they should stick to us.”
Microsoft agrees that leaderboards don’t tell you much. “What we see is that ‘benchmaxed’ models fall flat under real enterprise usage very quickly,” wrote Achint Srivastava, the engineering VP who led the project. He said only a few models, including his and Jev, hold up across all four dimensions that matter:
- Accuracy: picking the right option
- Reliability: behaving the same way on unseen, real-world inputs
- Speed: low enough latency for live production
- Cost: cheap enough to run at massive volume
🔮 Where This Goes in the Next 1-3 Years
This matters because it’s the clearest sign yet that AI stacks are splitting in two. Freddi describes a future where “System One” models (fast, intuitive decision models) work alongside “System Two” models (slow, deliberate frontier LLMs). That framing, borrowed from Daniel Kahneman’s work on human thinking, fits how agentic systems already run. Most steps in an agent loop are small choices, not essays.
Here’s what I expect next:
- Base decision models become a commodity. If a weekend project can compete, base model pricing will fall toward zero, just like it did for embeddings.
- Value moves to customization. The lasting businesses will own the feedback loop: customer data in, better calibration out.
- Hyperscalers bundle, specialists go deep. Microsoft and Amazon will fold decision models into their platforms. Startups will win on tuning tools and vertical expertise.
Treat Freddi’s market sizing with care. She puts System One at “maybe a $10 billion, $50 billion” opportunity, a guess from a founder with a pre-seed round to justify. Her strategic point holds up better. For OpenAI, which is chasing an IPO valuation she put at $1 trillion to $2 trillion, this is “just a side project.” For Sqwish, it’s everything.
🛠️ What Practitioners Should Do Now
- Audit your LLM calls. Find every place a frontier model is really making a multiple-choice decision. Those calls are probably costing you too much.
- Test on your own traffic. Swapping models is easy, so run candidates against real production data, not public benchmarks.
- Use the confidence scores. Send low-confidence calls to a frontier model or a human. That’s where the hybrid design pays off.
- Collect labeled outcomes now. Whoever owns clean decision data will get the most out of the tuning tools that are coming.
The frontier labs will keep getting the headlines, but a growing share of production AI now runs on fast, cheap decision models. Full details on the decision model race are available at TechCrunch AI.