The market for AI decision models is expanding rapidly, with customization emerging as the critical next frontier for differentiation. In October, more than 25% of requests on Vercel’s AI model router were directed to decision models rather than large language models (LLMs), a significant volume for a technology category that was introduced only a month prior.
Decision models are a specialized class of LLMs that do not generate output tokens. Instead, they are tuned to use general intelligence to quickly and cost-effectively select from preset options while registering their confidence levels. These models are particularly useful in production environments where rapid decision-making is prioritized over the slower, more expensive reasoning provided by frontier models.
TypeSafe, the creator of the Jev family of models, recently achieved a $7.5 billion valuation after launch, with its models handling over 1.5 billion queries last week through services like OpenRouter. The success of Jev has triggered a wave of competition; dozens of new decision models have been released, including one developed by an Amazon engineer in a single weekend, and Microsoft has announced the general availability of its Microsoft-Decision-1.1 model on Azure.
While many entrants offer basic functionality, data is proving to be the key to success. Diogo Almeida, CEO of TypeSafe, noted that calibrating the judgment of the Jev model using statistically adjusted synthetic data was the most challenging aspect of development. Meanwhile, UK-based startup Sqwish has secured $3 million in pre-seed funding to compete in this space. Sqwish’s cofounders, Ushnish Sengupta and Federica Freddi, pivoted from developing small LLMs to creating their decision model, kettei fat, leveraging their existing architectures.
Sqwwish differentiates itself through built-in post-training tools that allow customers to upload their own production data to improve model performance. “What makes us not a commodity… is that our customers 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.
Industry leaders agree that benchmark scores are less reliable indicators of real-world utility than continuous tuning and practical application. Achint Srivastava, the engineering VP at Microsoft who led the development of its decision model, stated that models optimized solely for benchmarks often underperform in enterprise environments. “Our testing underscores just how challenging it is to deliver accuracy, reliability, speed and cost efficiency in a single model,” Srivastava said, noting that Microsoft-Decision-1.1 performs strongly across these dimensions on unseen real-world cases.
For startups like Sqwish, the presence of hyperscalers and frontier labs represents both a challenge and an opportunity. Freddi predicted that future AI systems will integrate “System One” intelligence, represented by fast decision models, with “System Two” intelligence, represented by slower frontier LLMs. She argued that the potential market for System One intelligence could reach $10 billion to $50 billion, a niche that larger companies like OpenAI may view as a secondary focus compared to their trillion-dollar valuations.
System 1 AI makes sense for speed, but will enterprises trust automated choices without the slower reasoning of frontier models?
I am skeptical that generic models can replace nuanced human judgment. Customization helps, but is it really enough for complex business decisions?
The weekend project by an Amazon engineer shows how fast this space is moving. It feels like a gold rush right now.