Diogo Almeida, a former OpenAI researcher who co-authored the foundational InstructGPT paper, is betting that the future of enterprise AI isn’t about chatty bots but silent judges. His new startup, TypeSafe AI, has launched Jev, a model designed to skip text generation entirely. Instead of drafting emails or writing code, Jev evaluates predefined options and returns probabilities. It’s a shift from creative output to deterministic decision-making, aimed squarely at developers who need speed rather than flair.
The practical application is straightforward but powerful. Imagine a customer complaining about a double charge. Traditional LLMs might draft a polite apology, but Jev simply tags the issue as “payment” with a high probability of a refund request. This allows backend software to route the ticket to accounting automatically. TypeSafe claims this process takes between 70 and 500 milliseconds. That is significantly faster than standard language models, which often struggle to maintain sub-second latency when generating coherent paragraphs. By computing multiple outputs in parallel, Jev keeps response times flat even as developers add more classification questions.
Cost is the other major lever here. TypeSafe lists the price at $0.042 per million input tokens and charges nothing for outputs. Compare that to the hefty bills enterprises rack up running general-purpose models for simple sorting tasks. However, the technology isn’t magic. While TypeSafe markets Jev as hallucination-free, that label is misleading. The model cannot invent new answers outside its preset menu, but it can still pick the wrong option from that menu. A factually incorrect judgment within a constrained list is still an error, just a more structured one.
Skeptics will note that OpenAI already offers Structured Outputs, allowing standard models to return JSON or specific categories. The differentiator for Jev must be raw performance and reliability at scale, not just format. The company’s own benchmarks are limited, lacking independent verification or comparisons against the latest competitors like GPT-6 Astra. For now, Jev remains behind a waitlist. Developers will need to decide if the speed savings justify swapping a flexible generalist for a rigid specialist. If the accuracy holds up, it could redefine how we build AI agents, moving them from conversational partners to invisible infrastructure.
