Former OpenAI Researcher Diogo Almeida Launches Jev, a Hallucination-Free Decision Model, With $40 Million

On September 15, 2026, TypeSafe AI, founded by Diogo Almeida, who previously worked on ChatGPT development at OpenAI, released Jev, its first System One model, and simultaneously announced the completion of a $40 million funding round.

Factual Recap

According to the TypeSafe AI official blog, Jev is trained with a self-developed RLCD algorithm and uses a parallel sampler to output structured type values, with response times in the 70-500ms range. Official materials show that the model reaches intelligence levels comparable to existing LLMs on System One tasks while cutting costs by 40-400x, and that type errors are mathematically impossible.

Official comparisons note that existing LLMs, optimized via RLHF or RLVR, output strings that require subsequent parsing and validation, whereas Jev directly outputs type-safe structured values with calibrated probabilities; pricing is $0.042 per million input tokens, with output free of charge.

How It Works

Jev's parallel generation mechanism is fundamentally different from the token-by-token autoregressive generation of LLMs. Official descriptions show that Jev generates its entire output in a single query, and hardware-aware optimization delivers a 40-200x speed advantage at equivalent intelligence levels. The training objective of RLCD is calibrated decision-making—that is, providing probability outputs with confidence levels on tasks such as classification, scoring, and routing.

Official materials emphasize that Jev gives up string generation capability in exchange for being "physically incapable of hallucinating"; all possible outputs are predefined by the type system during training.

Industry Impact

Jev's release further segments the scenarios where existing LLMs apply. LLMs remain suited to open-ended tasks that require human involvement, such as chat and code generation, while Jev offers faster, cheaper options that can be embedded directly in code for verifiable decision steps. Official use cases include intelligent if statements in AI workflows, feature extraction from large datasets, and validation and guardrails for LLM outputs.

This division of labor could change how AI systems are integrated, letting developers use Jev for real-time decision nodes and LLMs for steps that require creativity or multi-step reasoning.

Strategic Assessment

[Analysis] If Jev's pricing and speed advantages can be sustained over the long term, they will put pressure on the share of existing LLMs in the decision-making API market; but the model's abandonment of general-purpose string generation means it cannot single-handedly handle complex application scenarios that require flexible output, and it may end up complementing LLMs rather than directly replacing them.

[Analysis] Whether non-LLM decision models should be included in the next round of evaluation systems depends on whether the industry accepts "structured output + calibrated probabilities" as an independent evaluation dimension; there is currently no public consensus.