TypeSafe launches Jev, a text-free AI for instant software decisions

For years, frontier AI labs have taught models to converse like humans. But when engineers try to embed those same models inside software for simple logic checks, they face high latency, steep API bills, and constant formatting errors.

TypeSafe AI is taking a completely different approach: stripping away text generation altogether.

The startup launched early access for Jev, a structured decision model designed specifically for software automation. Founded by Diogo Almeida, a former OpenAI researcher who helped develop the reinforcement learning methods behind ChatGPT, TypeSafe spent two years building what it calls a "System One" intuitive model.

No text, just parallel probabilities

Standard large language models generate text token by token. That sequential sampling makes them flexible, but it also creates latency bottlenecks and opens the door to hallucinations.

Jev abandons freeform strings. Instead, it relies on a training approach called Reinforcement Learning for Calibrated Decisions (RLCD). Developers define their expected schema upfront, whether it is a boolean, a classification choice, or a numerical score. Jev consumes unstructured text and evaluates all possible options in parallel, returning calibrated probabilities in a single query.

Because the output structure is strictly defined in advance, TypeSafe claims Jev cannot produce type errors or format hallucinations. Response times sit between 70 and 500 milliseconds, representing a 40x to 200x speedup over frontier models on comparable structured tasks.

Pricing is equally aggressive: input tokens cost $0.042 per million, while output tokens are entirely free.

Real-time sorting and a Doom demonstration

Engineers are already testing how Jev performs as an intelligent intermediary inside existing software stacks.

At Vercel, engineer Pranit Sharma replaced OpenAI's ChatGPT Luna 5.6 with Jev for command safety classification, reporting a 5x to 18x speed improvement alongside higher accuracy. Nikhil Mudholkar, CTO of Bryo AI, tested Jev for email triage; while Gemini was marginally more accurate, Jev was 10 to 20 times cheaper and provided dependable confidence scores.

To showcase real-time throughput, TypeSafe hooked Jev up to Doom, where the model made 10 game decisions per second at a cost of roughly $7 an hour. The team also showed it navigating Wikipedia links in Wikiracing challenges.

Armin Ronacher, CTO of Earendil, noted that Jev delegates the hallucination problem to software logic. If a decision comes back with 95% confidence, the app can automate it instantly; if it hits 50%, the system can flag it for human review. Ronacher also sees Jev acting as an affordable router to decide when to call frontier models.

Why it matters

The model takes its name from the Jevons paradox, which states that making a resource cheaper drastically increases how much people use it.

For developers, Jev turns AI from a slow chatbot interface into a dependable, lightweight control statement. Instead of wrestling with fragile JSON prompts and praying the LLM adheres to schema rules, engineers can drop Jev into background workflows for routing, validation, and real-time guardrails without wrecking user experience.

We think TypeSafe's bet on specialized System One models is remarkably clever. While major labs pour capital into slow, chain-of-thought reasoning systems, mundane software automation rarely needs deep philosophical reflection. Dropping token generation removes the overhead that makes LLMs painful to deploy in low-latency infrastructure.

Still, early adopters should keep limitations in mind. Predefined outputs mean Jev cannot tackle creative generation, and zero type errors do not guarantee zero mistakes. Since TypeSafe's internal benchmarks use averages from GPT-6 Astra and Claude Fable 5.1 as reference points, developers will need to measure calibrated accuracy on their own production data as waitlist access rolls out.