Decision models took over 25% of requests in Vercel's router a month after launching
Decision models are LLMs that don't write text. They quickly and cheaply pick one of several preset options and report how confident they are. In October they accounted for more than 25% of requests in Vercel's router, even though the class appeared only a month earlier. Microsoft launched its own model in Azure, and the startup Sqwish raised $3 million to fine-tune models on customer data. That is why these models are worth comparing on your own traffic rather than by benchmarks.
AI-processed from TechCrunch; edited by Hamidun News
If you build a product on LLMs, you almost certainly have tasks for which a large model is overkill. Deciding which category a ticket belongs to. Deciding whether to send a request to an expensive model or let a cheap one handle it. Figuring out whether something is spam. Over the past month, a separate class of models has emerged for tasks like these, and it is growing fast. According to TechCrunch, in October more than one in four requests in Vercel's AI router went not to regular LLMs but to decision models.
What is a decision model
It is an LLM that doesn't generate an answer in words. It is tuned to apply general "intelligence" to choose quickly and cheaply among predefined options, and to show how confident it is in the choice. Think of a duty administrator who doesn't write you an email but simply ticks the right box and adds: "90% sure." That suits production. When you need a decision in a fraction of a second and for pennies, a slow, expensive frontier model with long reasoning only gets in the way.
How fast this happened is striking. TypeSafe became a $7.5 billion company after releasing its Jev model family, and Jev still leads in usage, at least where the metrics are visible. For example, TypeSafe models received more than 1.5 billion requests via OpenRouter last week. But building such a model turned out to be easy. Dozens of analogs followed Jev, one Amazon engineer built his own over a weekend, and every hyperscaler now has a variant of its own. On Tuesday, Satya Nadella announced that Microsoft is launching Microsoft-Decision-1.1 straight into general availability in Azure.
Data matters more than architecture
If models are easy to build, how do you stand out? TypeSafe head Diogo Almeida says the hardest part of Jev was the fine-grained calibration of judgments on synthetic data with statistical correction. In other words, what matters is not only what the model chooses but also how honestly it rates its own confidence.
The British startup Sqwish, the maker of kettei fat, is betting on customization. Today it announced a $3 million pre-seed round. Co-founders Ushnish Sengupta and Federica Freddi met at Cambridge and worked on small LLMs. When Jev appeared, they realized their architectures could be reworked into a decision model. Sqwish's hook is built-in fine-tuning tools: a customer uploads its production data, and the model gets better specifically on that customer's tasks. According to the CEO, this is what sets the company apart from a commodity product. Customers get the same power that the authors used to train the models in-house.
Benchmarks are only an invitation
Comparing these models by tables is hard. kettei fat performed well on some benchmarks, but its developers themselves admit that you have to test it in practice. Swapping one model for another in a specific use case is usually easy. Sengupta calls a benchmark "an invitation to try," and says continuous fine-tuning should be what keeps a customer.
Achint Srivastava, the Microsoft vice president of engineering who led the project, made a similar point. In his words, "benchmaxxed" models quickly fail in real enterprise operation. Microsoft-Decision-1.1 balances cost, latency, intelligence and calibration. Together with Jev, the company claims, it is one of the few that is strong on all four axes, including on held-out tests close to real-world scenarios. Keep in mind that this is Microsoft's own claim, and the article contains no independent comparisons.
What this means for developers
Federica Freddi predicts that in the future "System 1" (fast decision models) will work in tandem with "System 2" (heavy frontier LLMs). Even in such a world, she believes, the "System 1" market could be worth $10–50 billion. For Sqwish it is the only thing the company does, whereas for OpenAI it is a side project. The forecast is, of course, a startup's forecast, not a guarantee.
The practical takeaway is simple. If part of your traffic is classification and routing, try moving it to a decision model: it is cheaper and faster. Choose by your own data, not by leaderboards. Swapping a model is easy, so run two or three candidates on real requests. Look not only at accuracy but also at confidence calibration, since that is what you will use to decide when to hand a task to a heavy model. And take a close look at vendors that offer fine-tuning on your data, because judging by the market, that is where the competition will be.
- Nov 18, 2026AI Finds Unknown V8 Vulnerability for $4.24 in 13 Minutes, After Failing for $350 in February
- Nov 18, 2026Anthropic: Claude with 413 note files earned 1 badge in Pokémon, but with a handful of files it earned 4
- Nov 18, 2026Opus 5.5 Tells 19 Safety Tests From 36 Ordinary Tasks Without a Single Error
- Nov 18, 2026GPT 6.2 Sol Ultrafast Delivers 150 Tokens/s but Burns Through Limits 6x Faster
Key Questions Answered
What is a decision model?
It is an LLM that doesn't generate an answer in words. It is tuned to apply general "intelligence" to choose quickly and cheaply among predefined options, and to show how confident it is in the choice. Think of a duty administrator who doesn't write you an email but simply ticks the right box and adds: "90% sure." That suits production. When you need a decision in a fraction of a second and for pennies, a slow, expensive frontier model with long reasoning only gets in the way. How fast this happened is striking.
What are the key details on Data matters more than architecture?
If models are easy to build, how do you stand out? TypeSafe head Diogo Almeida says the hardest part of Jev was the fine-grained calibration of judgments on synthetic data with statistical correction. In other words, what matters is not only what the model chooses but also how honestly it rates its own confidence. The British startup Sqwish, the maker of kettei fat, is betting on customization. Today it announced a $3 million pre-seed round.
What are the key details on Benchmarks are only an invitation?
Comparing these models by tables is hard. kettei fat performed well on some benchmarks, but its developers themselves admit that you have to test it in practice. Swapping one model for another in a specific use case is usually easy. Sengupta calls a benchmark "an invitation to try," and says continuous fine-tuning should be what keeps a customer. Achint Srivastava, the Microsoft vice president of engineering who led the project, made a similar point.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.