New AI model delivers fast

New AI model delivers fast - ai structured decisions — Jev processes inputs in parallel and returns organized outputs ready for immediate application without generating open-ended text.

TypeSafe AI has introduced Jev, a specialized model that delivers structured decisions instead of generating text. Unlike conventional large language models, Jev produces typed probabilities—Choice, Score, and Noul responses—paired with confidence values, enabling software to act directly on results that meet a predefined threshold. The system does not generate open-ended text; instead, it processes inputs in parallel and returns organized outputs ready for immediate application.

Operational Design

Users submit a state, either as a string or structured data, alongside typed questions. Jev evaluates all inputs in a single pass and returns answers with probability distributions. Pricing starts at $0.042 per million tokens for input, with output remaining free. The model supports a 32,000-token context window and achieves end-to-end latency between 70ms and 500ms. TypeSafe’s approach relies on Reinforcement Learning for Calibrated Decisions, diverging from traditional fine-tuning methods.

This design represents a deliberate shift in AI development, prioritizing speed, precision, and cost efficiency over generative flexibility. While models such as Gemini or GPT-5.6 remain strong in open-ended tasks, Jev targets high-speed decision-making scenarios where structured outputs are essential. The trade-off echoes earlier transitions, like moving from unconstrained text generation to structured outputs in tools such as LangChain, but emphasizes probabilistic certainty over creative ambiguity.

Early Adoption and Benchmarks

Vercel incorporated Jev into its AI Gateway shortly after launch, with nearly 13% of paid teams adopting it within 24 hours, a rate double that of GPT-5.6 on its first day. Netlify followed, and LangChain released a TypeSafeClassifier integration, including model routing and an AutoMode middleware to pre-filter tool calls. Five independent Elixir clients also appeared rapidly, reflecting strong developer interest in the model’s structured output capabilities.

Performance tests show Jev’s advantages. A Vercel engineer reported a safety classifier built with Jev ran five to 18 times faster than an equivalent large language model. Bryo AI’s chief technology officer noted that while Gemini achieved higher accuracy in email classification, it did so at 10 to 20 times the cost, calling Jev the only model providing genuine probability distributions. Earendil’s CTO observed that the model transfers some decision risk to users, who must assess whether a 50% probability warrants action, while model routing remains a practical solution.

Limitations and Developer Feedback

An analysis of 12,759 launch tweets by OpenChamber revealed median speedups of 7x (compared to a headline claim of 193.6x), cost savings of 30x, and latency averaging 76ms (with an upper quartile at 270ms). Developers on Reddit and Hacker News praised Jev for agent workloads, citing its 200ms to 300ms response times, though some warned that out-of-distribution behavior could still differ from traditional large language models.

A Hacker News contributor highlighted Jev’s inability to produce invalid types, though it could still return incorrect but syntactically valid responses. The documentation cautions against relying on the model for counting, arithmetic, or date comparisons, noting accuracy declines with large or noisy input states. TypeSafe recommends handling mathematical operations in code and using fixed versions like jev-1.13.0 instead of dynamic updates such as jev-latest. The quickstart guide includes instructions for API keys, SDKs, and a testing playground.