Decision models capture 25% of Vercel traffic as Microsoft enters the market
New AI models that output choices instead of text are dominating production workloads. Microsoft and startups race to optimize these fast, low-cost decision engines.
In October, more than one in four requests on Vercel’s AI model router were directed to decision models rather than traditional large language models. This surge represents a massive adoption rate for a technology category that was invented only a month prior. Major players like Microsoft and specialized startups are now rushing to define this new layer of artificial intelligence infrastructure.
What happened
Decision models are a distinct class of AI that do not generate output tokens like standard LLMs. Instead, they are tuned to use general intelligence to quickly choose between preset options while reporting a confidence level for that choice. This architecture is optimized for production environments where speed and cost efficiency outweigh the need for verbose reasoning or creative text generation. The rapid shift in usage patterns highlights a growing demand for lightweight, deterministic AI components in software stacks.
TypeSafe has emerged as an early leader in this space, reaching a valuation of $7.5 billion after releasing its Jev family of models. Metrics from OpenRouter show that TypeSafe’s models handled more than 1.5 billion queries in a single week, indicating dominant market share in visible channels. However, the barrier to entry appears low. Dozens of competing models have been released since Jev’s debut, including one built by an Amazon engineer in a single weekend. Every major hyperscaler now offers its own variant.
On Tuesday, Microsoft CEO Satya Nadella announced the immediate wide release of Microsoft-Decision-1.1 on Azure. Simultaneously, Sqwish, a U.K.-based startup, announced a $3 million pre-seed round. Sqwish’s model, called kettei fat, was developed by Cambridge graduates Ushnish Sengupta and Federica Freddi. They repurposed their existing small LLM architectures to compete in this new category, betting that customization and post-training tools will differentiate them from commodity offerings.
How it works
The core technical challenge in building effective decision models is not just raw intelligence, but calibration. TypeSafe CEO Diogo Almeida noted that the hardest part of creating Jev was finely calibrating the model’s judgment using statistically adjusted synthetic data. The model must accurately assess its own confidence to avoid making high-stakes errors in automated workflows. Without this calibration, a fast model is merely a fast source of hallucinations.
Sqwish approaches this problem by providing built-in post-training tools that allow customers to improve the model’s judgment using their own production data. This creates a feedback loop where the model adapts to specific business contexts over time. Rather than relying solely on static benchmarks, this method emphasizes continual tuning. Developers can swap models easily, but the value proposition shifts toward long-term performance improvement through proprietary data integration.
Microsoft-Decision-1.1 was engineered by balancing cost, latency, intelligence, and calibration. Achint Srivastava, the engineering VP who led the project, emphasized that models optimized strictly for benchmarks often fail in real enterprise usage. Microsoft’s testing focused on held-out benchmarks that reflect actual use cases, ensuring the model performs reliably across all four critical dimensions simultaneously.
Key details
- In October, over 25% of Vercel AI model router requests went to decision models.
- TypeSafe’s Jev models processed more than 1.5 billion queries in one week via OpenRouter.
- Microsoft released Microsoft-Decision-1.1 for wide use on Azure immediately.
- Sqwish raised $3 million in pre-seed funding for its kettei fat decision model.
- Decision models output preset choices and confidence scores, not text tokens.
- Calibration using synthetic and production data is cited as the primary technical hurdle.
Why it matters
For software engineers and technical founders, the rise of decision models signals a shift in how AI is integrated into products. Previously, developers might have used expensive frontier models for simple routing or classification tasks, paying for reasoning capabilities they did not need. Decision models offer a way to reduce latency and costs significantly for these high-volume, low-complexity tasks. This allows teams to reserve powerful, slower models for complex generative tasks while handling routine logic with specialized, efficient agents.
The competition between hyperscalers and startups also highlights the importance of data ownership and customization. While Microsoft and Amazon provide robust, general-purpose decision models, startups like Sqwish argue that generic solutions will become commodities. The ability to fine-tune a model on private production data offers a defensive moat for companies building proprietary AI features. As Federica Freddi noted, this "System One" intelligence represents a multi-billion dollar opportunity distinct from the trillion-dollar valuations of general-purpose AI labs.
What you can do
- Audit your current AI usage to identify high-volume, low-complexity tasks suitable for decision models.
- Test Microsoft-Decision-1.1 or TypeSafe’s Jev models against your current LLM routers for cost savings.
- Evaluate whether your application requires custom calibration via post-training tools offered by startups like Sqwish.
- Monitor confidence scores from decision models to set appropriate thresholds for human-in-the-loop interventions.
- Prepare synthetic datasets that reflect your specific business logic to improve model calibration.
- Consider architecting systems where "System One" decision models handle routing before passing complex cases to "System Two" generative models.



