TypeSafe AI exits stealth with Jev, a model that outputs decisions, not tokens

TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, emerged from stealth with Jev — a transformer that generates no tokens and instead returns calibrated probabilities or typed values from a fixed schema. The company claims 70–500ms latency, $0.042 per million input tokens and up to 193.6x speedup over LLMs, and has raised $40M.

• Jev returns a value from up to 255 options or a 0–1 probability, generating no tokens
• Latency 70–500ms; demo showed 0.114s vs 8.566s for GPT-5.6 Terra
• Pricing $0.042 per million input tokens, output tokens are free
• Trained on synthetic data via RLCD, optimizing calibration instead of RLHF