TypeSafe AI Unveils Jev Model 18x Faster for Automation as Demand Briefly Overloads API
Updated
Updated · TechCrunch · Sep 18
TypeSafe AI Unveils Jev Model 18x Faster for Automation as Demand Briefly Overloads API
3 articles · Updated · TechCrunch · Sep 18
Summary
TypeSafe AI said Jev outputs probabilities rather than text, aiming software automation at calibrated decisions that users predefine instead of open-ended language generation.
5 to 18 times faster results at Vercel and 10 to 20 times lower costs than Gemini in one email-classification test helped drive early developer interest, which briefly knocked the startup’s API offline.
Jev’s fixed-output design makes output tokens free, meters input by the billion rather than the million, and avoids hallucinations by limiting answers to preset choices with confidence scores.
Diogo Almeida, who left OpenAI two years ago, argues LLMs are optimized for human language rather than computer workflows; TypeSafe says Jev is trained on synthetic data using “reinforcement learning from calibrated decisions.”
TypeSafe sees Jev replacing LLMs in some classifiers while also monitoring agent traces, blocking jailbreaks and routing workloads cheaply enough for broader, distributed use of AI in software.