TypeSafe AI launches Jev: A fast model built for structured output, not chat
TypeSafe AI just released Jev, a model that abandons string generation to output type-safe decisions up to 200x faster than normal LLMs.
Models · Source: Hacker News
What happened
Diogo Almeida and the team at TypeSafe AI just ended a two-year stealth period. They released Jev, their first System One model. It is built specifically for software automation, not chatting with humans. Almeida previously worked at OpenAI on the research that led to ChatGPT but realized chat models were missing the reliability needed for true automation.
Jev takes unstructured data as input and returns typed, probabilistic decisions. It completely drops the ability to generate strings. This means it cannot hallucinate type errors or go off the rails with unexpected text. The model uses a parallel sampler instead of generating tokens sequentially. It is trained using a new method called Reinforcement Learning for Calibrated Decisions.
The performance claims are massive and backed by transparent pricing. TypeSafe says Jev is up to 200 times faster than frontier models for specific tasks. Response times sit between 70 and 500 milliseconds. Input tokens cost just 42 dollars per billion, and output tokens are completely free.
Key facts
- $0.042 — Cost per million input tokens for Jev
- 70ms-500ms — End-to-end response time for TypeSafe models
- 193.6x — Claimed speedup over frontier models on specific workflows
- 444.6x — Claimed cost reduction compared to standard LLM workflows
Why it matters
Builders waste too much time parsing JSON and guarding against hallucinations. When you use a standard language model for a routing or classification task, you pay for the overhead of sequential text generation. Jev treats artificial intelligence like a pure function call. You define the structure in advance. The model simply returns the structured values with calibrated confidence scores. Higher confidence actually means higher accuracy, making the model highly consistent for similar inputs.
The second-order effect is real-time AI integration across the board. At 100-millisecond speeds, you can put AI directly into user-facing application loops without ruining the experience. The TypeSafe team even built a Doom bot making ten queries a second for just seven dollars an hour. When intelligence gets this cheap and fast, we will stop building fragile wrappers. We will start using AI as standard control logic in traditional software architectures.
For builders
Stop paying for output tokens in automation
Jev charges a fraction of a cent per million input tokens and makes output tokens entirely free. If you run high-volume classification or data extraction, you save massive amounts of money. Traditional frontier model providers lose out on these pure routing workloads.
Build real-time user experiences with AI
End-to-end response times under 500 milliseconds mean you can use AI in the critical path of your user interface. You no longer have to hide latency behind long loading spinners. Engineers building interactive tools get a massive performance edge over competitors using slow chat models.
Eliminate prompt engineering for JSON outputs
You do not need to beg the model to return valid JSON anymore. The schema is guaranteed by the architecture itself. Builders save countless hours of debugging brittle prompts and writing complex fallback parsing logic.
My take
I have been saying for years that chat is a dead end for real automation. TypeSafe is finally building the AI primitive software engineers actually need. Forcing a language model to act like a deterministic function was always a hack, and Jev proves that native structured output is the future of AI infrastructure.
Original reporting: Hacker News. This is my rewrite and opinion.