Introduction

Artificial intelligence chatbots are impressive, but they often invent facts and respond slowly. A new breed of model called a decision model aims to solve both problems by choosing from preset answers instead of generating free-form text. TypeSafe AI kicked off the trend with Jev, and now major cloud providers and AI labs are rushing to release their own versions.

What Happened

On September 15, TypeSafe AI unveiled Jev, which it describes as a System One model. The company claims the system delivers responses in as little as 70 milliseconds, costs far less than conventional large language models, and cannot hallucinate because it only selects from user-defined options. The startup recently secured $870 million in funding at a $7.5 billion valuation.

Soon after, Amazon, Cloudflare, Microsoft, and OpenAI introduced decision models inspired by Jev. Independent developers and researchers have published dozens more. Cloudflare, for example, released Clef and Claric, with Clef available as open weights on Hugging Face and built on Alibaba's Qwen model.

Research cited by Carnegie Mellon professor Ramayya Krishnan recorded Jev's median response time at 150 milliseconds on decision benchmarks, versus 1.89 seconds for OpenAI's GPT-6 Astra. TypeSafe's pricing sets input tokens at $0.042 per million and makes output tokens free. Competitive LLM pricing can reach $10 per million input tokens and $50 per million output tokens for comparable systems.

Why This Matters

Traditional language models generate text one token at a time, which creates opportunities for errors and requires substantial computing power. Decision models use a single forward pass to evaluate inputs and return one structured answer, such as selecting option A, B, or C with an associated probability. This design makes them ideal for high-volume tasks like content moderation, spam detection, phishing identification, and prompt-injection defense.

Cloudflare product manager Michelle Chen noted that teams previously needed custom data and machine-learning engineers to build classifiers for each job. A general-purpose decision model can be applied across many domains without that overhead.

Key Takeaways

  • Decision models choose from predefined answers rather than generating open-ended text, which eliminates certain types of hallucination.
  • Jev and its imitators promise response times measured in milliseconds rather than seconds.
  • Input token pricing for Jev is a fraction of what major LLM providers charge, and output tokens are free.
  • These models still struggle with math, logic puzzles, and specialized domains such as medicine, finance, and law.
  • Hybrid setups that route uncertain decisions to a larger LLM can improve accuracy while reducing cost.

Conclusion

Decision models are not a complete replacement for large language models. Research from Carnegie Mellon and Fudan University shows they trail general LLMs on tasks requiring deep subject expertise. However, they offer a practical way to handle routine judgments quickly and cheaply. Many organizations are likely to adopt a hybrid approach, using fast decision models for simple calls and passing difficult cases to more powerful, slower systems.