TypeSafe AI

Artificial Intelligence · AI Infrastructure

Launch

TypeSafe AI launches Jev, a frontier model built for machines, not chat

Removing text generation entirely lets software call AI like a typed function, closing the reliability gap that stalls automation.

• TypeSafe AI emerged from stealth on September 15, 2026, launching Jev, its first 'System One' model, alongside a $40 million Series Seed led by DCVC.
• The company was founded by Diogo Almeida, a former OpenAI researcher who co-invented the RLHF techniques behind InstructGPT and ChatGPT, alongside Erik Gafni and Sasha Sheng.
• Instead of chat responses, Jev takes in program state and typed questions and returns structured probabilistic answers, like {"billing": 0.08, "technical": 0.85}, so software can act on them directly without parsing text.
• TypeSafe claims Jev responds in 70-500 milliseconds, 20-200x faster than comparable LLMs, and is priced at $0.042 per million input tokens with free output, versus $2-$12 per million for rivals like GPT-5.6 Terra.
• In one of the first independent tests, Every's evals team had Jev make 777 judgments across 37 documents in under a second for about a quarter of a cent, catching six of seven deliberately planted errors versus seven of seven for Claude Fable 5.1.
• TypeSafe is betting that most future AI usage will be software calling software rather than humans chatting with assistants, a wager that automation has stalled because models were optimized for conversation, not machine-to-machine reliability.