I have been thinking about this since I first read the TechCrunch headline: a company called TypeSafe AI just raised $870 million at a $7.5 billion valuation — and they launched their product three weeks ago. Three weeks. That is not a typo. That is the fastest I have ever seen a startup go from public debut to decacorn status.
The product is called Jev, and it is not a large language model. It does not write essays, generate code, or chat with you. It does something that, in the current AI landscape, is almost radical: it makes structured decisions. It takes unstructured input and outputs probabilities — calibrated decisions that software can use directly. No text generation. No token-by-token autoregressive decoding. Just fast, cheap, reliable decisions.
I spent the last two days reading everything I could about Jev — the company blog, the technical docs, the TechCrunch coverage, the Bloomberg piece. And I think this might be the most important AI story that nobody is talking about, because it challenges an assumption most of us have quietly accepted: that the future of AI is chat.
The Language Problem Nobody Wanted to Talk About
Here is the thing about LLMs that we do not discuss enough: they are built to generate human language. Every single token they produce is optimized for one thing — sounding plausible to a person. That is their entire training objective. RLHF, RLVR, instruction tuning — all of it is aimed at making the model better at communicating with humans.
But computers do not communicate in paragraphs. They communicate in structured data. JSON. Probabilities. Type-safe values. When you want an AI to make a decision — classify this email, route this support ticket, approve this loan application — you do not need it to write you a nice explanation. You need it to output a clean, typed, probabilistic answer that your code can consume without parsing, without hoping the format is correct, without worrying the model hallucinated a field.
That is what Jev does. It gives up string generation entirely and replaces it with a new architecture designed for one thing: calibrated decisions.
What “Non-Text AI” Actually Means
Let me break down what TypeSafe is claiming, because the technical details matter here.
Jev is built on a transformer architecture — same foundational idea as GPT, Claude, Gemini. But instead of generating tokens one at a time (autoregressive decoding), Jev outputs all probabilities in parallel. It is a single-pass model. You feed it unstructured state — text, JSON, structured fields — and it outputs a type-safe structured value with calibrated probabilities and confidence scores.
The training method is called Reinforcement Learning for Calibrated Decisions (RLCD) — a new approach distinct from RLHF or RLVR. The objective is not human preference or verifiable rewards. It is epistemic honesty: the model should know what it does not know. If it can do a task 95% of the time, it should tell you when it is in the 5%.
The numbers are staggering, if you take them at face value:
- Speed: 70ms to 500ms end-to-end response time, compared to 3 to 329 seconds for frontier LLMs. That is 40x to 200x faster.
- Cost: $0.042 per million input tokens, output tokens free. Compared to LLMs where output tokens cost ~5x more than input, and input runs $0.20 to $10 per MTok.
- Accuracy: The company claims 193.6x faster and 444.6x cheaper on their benchmark workflows — though they admit these are “on the higher end of real world gains.”
I am not going to pretend these numbers are independently verified. And as the fired OpenAI safety researchers warned, the AI industry has a credibility problem when it comes to self-assessment. They are company benchmarks on company-chosen workflows. But even if the real-world advantage is only 10x — that is still transformative for any organization running AI at scale.
The Fortune 500 Test
Here is what convinced me this is more than vaporware: a third of Fortune 500 companies are already using Jev. TypeSafe declined to name them, but the claim is specific enough that it would be falsifiable if it were completely made up. These are companies with legal teams, procurement processes, and zero tolerance for hype cycles.
And the adoption curve: one million users in a matter of days after the September 15 release. That is not a gradual enterprise sales cycle. That is developers inside these companies trying Jev, finding it works, and spreading it organically.
The co-founder story also matters. Diogo Almeida was a researcher at OpenAI — he helped build the methods that made language models useful at following instructions, the research behind ChatGPT. He left because he saw something missing. Sasha Sheng was a Meta research engineer. Erik Gafni is an engineer and entrepreneur. These are not hype-driven founders chasing a trend. They spent two years in stealth before launching.
The Jevons Paradox Connection
I love that they named the model after William Stanley Jevons, the 19th-century economist who observed that when steam-engine efficiency improved, coal consumption went up — not down. The Jevons paradox: efficiency gains increase demand rather than decrease it.
TypeSafe is betting the same thing will happen with AI. Every order of magnitude drop in the cost of intelligence unlocks orders of magnitude more use cases. It is the same dynamic that governs the entire AI industry economics, something I explored in my analysis of OpenAI’s $20 billion revenue gap. If you can make a calibrated decision for $0.000042 per input token instead of $1.50, you suddenly automate tasks that were never economically viable before.
This is the same dynamic that happened with cloud computing. AWS did not just make servers cheaper — it made possible business models that never existed before. Uber, Airbnb, Netflix streaming — all of them were economically impossible at the cost structure of traditional infrastructure. Jev is betting it can do the same for AI decisions.
Why I Think This Matters for Developers
Let me get practical for a moment, because I think this has real implications for how we build software.
Most of us working with AI today have dealt with the same frustration: you want the model to make a decision, but it keeps trying to have a conversation with you. You ask it to classify an email, and it writes you a paragraph explaining its reasoning. You parse that paragraph, hoping the format is correct, hoping it did not hallucinate a category. You write validation logic. You add retry logic. You add fallback logic for when the model goes off the rails.
That is the current state of AI integration. It is brittle. It is expensive. It is slow. And as I discussed in my piece on Tony Fadell’s critique of AI gadgets, the first wave of AI products failed precisely because they solved no real problems.
Now imagine a model that outputs a clean JSON object with the classification, the confidence score, and a calibrated probability. This is the kind of AI-powered automation that tools like Anthropic’s OSS Scanner are already bringing to software development. — in 70 milliseconds, for a fraction of a cent. You do not need to parse anything. You do not need to validate the format. The output is type-safe. It is guaranteed to conform to the schema you defined.
That is not a marginal improvement. That is a different category of tool. It is the difference between trying to have a conversation with a very smart person who keeps going off on tangents, and having a function that just returns the answer.
The Skeptic in Me
I would be doing you a disservice if I did not flag the reasons for caution here.
The benchmarks are company-run. The 193.6x speedup and 444.6x cost advantage are from TypeSafe’s own evaluation framework. The company admits these are “on the higher end of real world gains.” Independent third-party validation does not exist yet.
The workflows may be cherry-picked. The benchmark workflows were designed by TypeSafe’s own model capabilities team. The company says they were “not deliberately chosen to make our model look good,” but some bias could exist. The reference model used for comparison is the average of GPT-6 Astra and Fable 5.1 — which biases against OpenAI and Anthropic models.
The pricing may be subsidized. TypeSafe says as much: “We can’t prove it isn’t subsidized; we’ll need the long-term to prove the sustainability of our pricing (which we expect to go down, not up).” That is an honest statement, but it is also a red flag. $0.042 per MTok is absurdly cheap. If it is real, it is disruptive. If it is a loss-leader, it will not last.
The “cannot hallucinate” claim is strong. TypeSafe says Jev is mathematically impossible to hallucinate because outputs are type-safe. That is a structural claim, not an empirical one. I want to see it tested in production before I believe it completely.
What I Am Watching
Here is what I will be tracking over the next few months:
- Independent benchmarks. Once researchers outside TypeSafe start testing Jev on standard tasks, we will have a real picture of its capabilities and limitations.
- Production deployments. Are the Fortune 500 companies actually using this in production, or just kicking the tires? The answer will show up in case studies, job postings, and conference talks.
- Pricing sustainability. If TypeSafe can maintain $0.042 per MTok while scaling, the Jevons paradox kicks in and the market explodes. If they have to raise prices, the story changes.
- The competitive response. OpenAI, Anthropic, and Google are not going to sit still. If Jev’s approach is as good as claimed, we will see copycat architectures within a year.
The Bigger Picture
Step back for a moment and think about what this means for the AI industry.
For the last four years, the story has been: bigger models, more parameters, more training data, more compute. The entire industry has been scaling along one axis — language capability. And it has been revolutionary. ChatGPT changed how we think about computers. Claude changed how we think about coding assistants. Gemini changed how we think about multimodal AI.
But there is a ceiling to that approach. LLMs are expensive. They are slow. They hallucinate. They require human oversight. They are built for humans, not for software.
Jev represents a different bet: that the next wave of AI value is not in generating better text, but in making better decisions. That the real market is not chatbots — it is automation. Structured outputs, calibrated probabilities, type-safe values that software can consume without friction.
If that bet is right, the implications are enormous. Every system that currently uses brittle, expensive, slow AI for decision-making — fraud detection, content moderation, customer routing, document classification — becomes an order of magnitude cheaper and faster. That is not just a product launch. That is a platform shift. And it reminds me of the quality control crisis at Google, where AI-generated slop froze an entire bug bounty program.
And if the bet is wrong? Then TypeSafe AI just raised $870 million at a $7.5 billion valuation for a very expensive chatbot that cannot chat. Which would be one of the most entertaining startup failures in history.
Where I Stand
I am not going to tell you to go all-in on Jev. I am not going to tell you it is the future of AI. What I am going to tell you is this: watch this space.
The last time I saw a startup grow this fast, it was a little company called OpenAI, and they launched something called ChatGPT in November 2022. Three weeks later, it had 100 million users. By January 2023, it was the fastest-growing consumer application in history.
Jev is not ChatGPT. It is not trying to be. It is something different — something that might matter more to the people who build software than to the people who chat with it. And that is why I think it is worth your attention.
I am not saying the future is non-text AI. I am saying the future is probably both — language for humans, decisions for software — and the companies that figure out how to bridge that gap will define the next era of computing. TypeSafe is the first company to bet everything on that thesis. Whether they win or lose, they are going to make the rest of the industry think differently about what AI is for.
And that, more than any benchmark or valuation, is why I think this story matters.