Every few months an AI launch makes me stop scrolling. This week it was Jev, a new model from a San Francisco startup called TypeSafe AI, which just came out of two years in stealth. The founder is Diogo Almeida — a former OpenAI researcher who worked on the instruction-following methods that made ChatGPT actually usable, which is why the press keeps calling him a ChatGPT co-inventor. But here's the twist: his new model refuses to chat. Not one word.
01.What is Jev, exactly?
Think about how you use ChatGPT or Gemini today. You type a question, the model generates text, and either you read it or a program parses it. Jev throws away that last part. Instead of producing natural language, it takes an unstructured state — a support ticket, a game screen, a webpage — and returns typed probabilistic decisions: structured values with confidence scores that other software can act on directly. As Almeida put it, "unstructured state in, typed probabilistic decisions out."
If you've ever written code, this is the AI version of type safety. The output can't be a beautifully written paragraph with a wrong number hiding inside it, because the output isn't a paragraph at all. It's a schema — a fixed set of fields the model must fill. If you're still building your basics, our GPT-6 Astra beginner explainer covers how today's frontier models work before you go comparing newcomers like Jev against them.
02.Speed: 70 milliseconds instead of minutes
Because Jev doesn't generate text token by token, it doesn't inherit the agonising wait of a chatty model. TypeSafe reports end-to-end response times of 70–500 milliseconds, compared to anywhere from 3 seconds to 329 seconds for conversational frontier models. On fixed workflow benchmarks, the company claims its model ran up to 193.6 times faster than Claude Sonnet 5 and roughly 444 times cheaper than Opus 5 on classification tasks.
Honestly, what surprised me most wasn't the speed itself — it was the demo. They ran a bot playing the classic game Doom in real time, querying the model 10 times per second at a cost of about $7 an hour. That's the difference between AI you wait for and AI that keeps up with the world. It's a sharp contrast to the multimodal voice models we saw Google ship with Gemini 3.8 Live — the industry is simultaneously racing toward richer conversation and, with Jev, toward no conversation at all.
03.The pricing made me do a double take
Jev costs $0.042 per million input tokens — about ₹3.5 — and output is free, because there's no token-by-token generation to bill for. Compare that with a leading conversational model like GPT-5.6 Terra at $2 per million input tokens and $12 per million output tokens. If you're classifying millions of customer tickets or product listings, that's not a discount, that's a different universe.
This is exactly the kind of arithmetic we drill in the cost-optimization chapters of our GPT-6 Astra — Zero to Hero Guide — knowing when a small, specialised model beats a big conversational one is worth more than any single prompt trick.
04.How it works: RLCD, not RLHF
Under the hood, Jev is trained with what TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD). Instead of being rewarded for pleasing humans with fluent answers, it's rewarded for probabilities that match reality — when Jev says 90% confidence, it's right roughly 90% of the time. A hardware-aware parallel sampler evaluates all possible structured values at once (up to 255 discrete options) rather than writing them out one by one. The name itself is a nod to 19th-century economist William Stanley Jevons, who showed that making a resource cheaper often makes us use more of it, not less. Fitting, for a model that's this cheap.
05.Where India's builders should pay attention
The realistic use cases are unglamorous but lucrative: routing customer support tickets, moderating content, classifying huge datasets, verifying what another AI produced, making split-second decisions in logistics and fintech. Now think of the Indian context — a D2C brand in Surat sorting 50,000 support tickets a day, a Bengaluru fintech doing real-time fraud checks, an outsourcing firm tagging lakhs of documents. Latency and per-token cost are exactly the two numbers that decide whether their AI unit is a cost centre or a profit centre.
One caveat worth stating plainly: the 193x and "hallucination-free" claims come from TypeSafe's own evaluations, and being "hallucination-free" is easier when you never generate language in the first place. Jev entered early access on September 16, so independent numbers will take time.
Key Takeaways
- Jev is a "chatless" AI from TypeSafe AI, founded by ChatGPT co-inventor Diogo Almeida
- It returns typed, calibrated decisions in 70–500 ms instead of generating text
- About ₹3.5 ($0.042) per million input tokens, free output — versus $2/$12 for GPT-5.6 Terra
- Best suited for routing, classification, verification and real-time decisions — not conversation
- $40M seed round led by DCVC; early access opened September 16, 2026
Frequently Asked Questions
What is Jev from TypeSafe AI?
Jev is a “chatless” AI model from TypeSafe AI, a startup led by ChatGPT co-inventor Diogo Almeida. Instead of generating conversation, it returns typed, probabilistic decisions for software in 70–500 milliseconds.
How fast and cheap is Jev compared to other models?
TypeSafe AI claims up to 193x faster performance than Claude Sonnet 5 on workflow benchmarks. Input costs $0.042 per million tokens (roughly ₹3.5) with output completely free, versus $2/$12 for GPT-5.6 Terra.
What is Jev best used for?
Jev is best suited for routing, classification, verification and real-time decisions — not conversation. It is trained using RLCD (reinforcement learning from calibrated decisions) rather than RLHF, and early access opened September 16, 2026, alongside a $40M seed round led by DCVC.
What This Means for You
If you're a student or developer, this is a signal worth reading: the frontier is splitting into "models that talk to humans" and "models that talk to machines," and the second category is where a lot of Indian engineering jobs and freelance gigs will live. If you run a business, the right question is no longer "which chatbot should I subscribe to?" but "which decisions in my company can a machine make in 100 milliseconds for pennies?" That question — and the workflow thinking behind it — is what separates people who watch AI news from people who profit from it. If you want that thinking laid out step by step, from prompting to agents to cost control, our GPT-6 Astra — Zero to Hero Guide is a good place to start.