The Most Useful AI Model This Month Cannot Write a Sentence

Prod, not God

TypeSafe's tagline is "Build Prod, Not God." Diogo Almeida, the company's CEO, closed his launch video with the same line.

It is a jab at where the biggest AI labs are spending their energy, and I think it lands.

Anthropic is chasing the biggest version of the idea. Its CEO has written about AI compressing decades of scientific progress into a few years, and most of the company's messaging is about building extremely capable, general models safely. Call it AGI or call it building God, the target is a model that can do almost anything.

OpenAI has gone the other way. ChatGPT has become a consumer product that wants a hand in everything: search, shopping, browsing, coding, video, and eventually hardware.

Both paths produce impressive products, and I use them. Neither is aimed at the operations manager who needs 20,000 invoices sorted correctly by Friday.

Almeida helped build the chat era. He spent about four years at OpenAI working on RLHF, InstructGPT, ChatGPT, and GPT-4. In a podcast after the launch, he asked how AI can be smart enough to work on the hardest problems in math and still fail to automate basic, repetitive office work.

TypeSafe's manifesto gives its answer. Today's models are already smart enough to create a lot of economic value. They are hard to build on because they were trained to please people in conversation.

Jev is TypeSafe's attempt to fix that.

What Jev is

The easiest way I can describe Jev is as a smart "if" statement.

Normal software is good at simple rules. If the order total is above $100, apply free shipping. If the invoice is past due, send a reminder.

It falls apart when the rule needs judgment. Is this customer upset? Is this email about billing or a technical problem? Does this lead fit our ideal customer? Is this resume relevant to the role?

Businesses answer these questions all day, usually by having a person read and decide. Jev is built for them.

You give it some information, which TypeSafe calls the "state." That could be a support ticket, an email, a form submission, or a row from a spreadsheet. Then you ask it questions, and you decide in advance what the possible answers are.

Jev answers in one of three shapes:

• Yes or no, with a probability. "Is the customer asking for a refund?" 93%.
• One option from a list you define. "Which team should handle this: billing, technical, sales, or other?"
• A position on a scale you describe. "How frustrated is this customer, from calm to very angry?"

Every answer comes back with a confidence level. Jev says "billing, and I am fairly sure," or "billing, but this could also be technical." A chat model gives you an answer and leaves you to guess how much to trust it.

In the official announcement, TypeSafe calls Jev a "frontier-intelligence function call." Messy information goes in and clean decisions come out.

Why businesses need something like this

Most businesses I talk to do not need an AI that writes a beautiful paragraph. They need something that can look at the 400 emails that came in overnight and correctly pick the 12 that need a human today.

Of course, it is a bit more complicated than that.

Chat models can do that, but they are a poor fit for the job in three ways.

They are slow. A frontier model can take several seconds to respond, sometimes minutes. That works when a person is waiting for an answer. It breaks down when the decision is one step inside an automated process that runs thousands of times a day.

They are expensive. You pay for a model that can write code, reason through hard problems, and hold long conversations, and you get one word back.

They are unpredictable. You ask for "billing," and sometimes you get "Billing Department," or an explanation, or a format your system cannot read. Engineers end up writing extra code to clean up and double-check the output.

Confidence is the biggest problem. A chat model that is right 95% of the time is impressive. If it cannot tell you which 5% it might be wrong about, someone still has to check everything, and you have not automated much.

Jev is designed around that gap.

Why the idea makes sense

The label "System One" comes from Daniel Kahneman's Thinking, Fast and Slow. System 1 is fast, intuitive judgment, like glancing at an email and knowing it is spam. System 2 is slow, deliberate reasoning, like working through a complicated contract.

Large language models, especially reasoning models, are System 2 tools. They think, write, and explain.

Most decisions inside a business are System 1 problems: categorize this, flag that, route this ticket, approve or escalate. Using a reasoning model for them is like asking your most senior analyst to sort the mail.

The name Jev comes from William Stanley Jevons. He was the economist who noticed that as steam engines became more efficient, people used more coal, because cheaper power created new uses.

TypeSafe is betting intelligence will follow the same curve. When a decision costs a fraction of a cent and comes back in a fraction of a second, you start putting AI into places you would never have considered.

I think that bet is right. Cheaper decisions change which projects are worth attempting, and that matters more than the savings on existing work.

What it is good at

Jev does best on quick, repeated judgments with a fixed set of possible answers. TypeSafe's own material and the projects people built in the first two weeks point to five areas.

Sorting and routing. Support tickets, inbound emails, leads, and requests, each sent to the right team, priority, and next step.

Scoring against a rubric. Resumes against a role, leads against your ideal customer, or vendors against your criteria.

Checking other AI. An LLM writes a summary, and Jev checks whether each claim is supported by the source. An AI agent wants to take an action, and Jev first judges whether that action looks risky.

Labeling at scale. One developer ran his resume against 6,245 Y Combinator companies in 25 seconds for $0.37. Another analyzed 724 ads from 37 brands in about 40 seconds for roughly $0.09.

Real-time decisions. TypeSafe showed Jev playing Doom at about ten decisions per second, which it estimated at roughly $7 an hour. It also showed Jev playing Wikiracing, a game where you get from one Wikipedia page to another using only links, choosing among hundreds of options at each step. People have since had it play Pac-Man and Tetris, drive browser actions, and run simple simulations.

There is one early production example. Metaview, a recruiting platform, reported that it added Jev across its agents over a weekend. It said candidate searches went from minutes to seconds, with the same accuracy and a lower cost per search. That number comes from the company itself, and nobody has measured it independently yet.

What it can help with day to day

Customer service. A customer service team gets a ticket. Before anyone reads it, the system knows it is a billing issue, the customer is frustrated, they want a refund, and they mentioned canceling. It lands in the right queue with the right priority.

Sales. A sales team gets a form submission. The system scores how well it fits the ideal customer, whether it mentions a budget or timeline, and whether it is spam. Good leads reach a rep in seconds instead of the next morning.

Finance. A finance team receives invoices by email. The system decides which messages are invoices, which vendor category each belongs to, and whether anything looks unusual enough to review.

Hiring. A hiring team receives 800 applications. The system scores each one against written criteria and flags the uncertain ones for a person instead of silently rejecting them.

E-commerce. An e-commerce brand gets product reviews. The system separates shipping complaints from product complaints and notices when one issue starts showing up more often.

Each of these tasks eats hours of someone's week, and none of them needs a model that can write poetry.

What this means for a business

The cost changes the math

Jev costs $0.042 per million input tokens, and TypeSafe does not charge for output. Frontier LLMs typically charge $0.20 to $10 per million input tokens, and more for output.

Take a larger company making 1 million small AI decisions a day across support, sales, and operations. That is about 30 million decisions a month. Assume each decision needs a few paragraphs of context.

On a mid-tier frontier model priced at $2 per million input tokens and $10 per million output tokens, each decision costs roughly $0.0018. That comes to about $54,000 a month. The same volume on Jev comes to roughly $630 a month.

Those are list prices. Batching and caching discounts narrow the gap, and you still pay for LLM calls wherever you need writing. Even a fraction of that difference changes what a company can afford to automate.

Speed opens up new places for AI

TypeSafe reports responses in 70 to 500 milliseconds. Frontier models take anywhere from a few seconds to several minutes. Independent testers have reported similar numbers for Jev, though results depend on location and setup.

At that speed, a decision can happen while a customer is still on the page, while a form is being submitted, or in the middle of an agent's task. It no longer has to wait for an overnight batch job.

Confidence becomes a policy you can manage

With Jev, you set simple rules. If the model is very confident, act automatically. If it is somewhat confident, ask a person to confirm. If it is unsure, send the case to a human.

Risk tolerance moves out of a hidden prompt and into a number you can see, discuss, and adjust. A refund under $50 might need 80% confidence. Closing an account might need 95% and a human sign-off anyway.

Operations teams already think about risk this way. Jev lets them apply the same thinking to AI decisions, and I expect this to be the part business leaders care about most.

When to use Jev instead of OpenAI or Anthropic models

I don't find "Jev versus ChatGPT" a useful comparison. Jev does a different job, and it works best alongside GPT or Claude.

The way I think about it, Jev decides and the LLM writes.

Use Jev when:

• the answer is one of a known set of options
• you make the same kind of decision many times
• speed matters because a person or a system is waiting
• you need to know how confident the answer is before acting
• running a large model for every small decision is getting expensive

Use an OpenAI, Anthropic, or other LLM when:

• you need text, such as a reply, report, summary, or email
• the task needs reasoning or planning across many steps
• the answer cannot be defined in advance
• you need to work with images, audio, or video

Picture a stream of customer messages.

• When Jev decides a message is a simple order-status question, the system looks up the order with no LLM involved.
• When a complaint is complicated, Jev passes it to Claude or GPT with the right context to write a careful reply.
• When a customer is angry and threatening to leave, the message goes straight to a person.

Jev can also decide which LLM gets the work. Simple requests go to a cheaper, faster model, and hard ones go to the expensive model. That routing alone could cut a company's AI bill noticeably.

What it is not good at

A lot of the coverage has been too generous here. Jev has real limits, and TypeSafe documents many of them itself.

It cannot write. It produces no replies, summaries, or explanations. If the output needs to be words, use a different model.

It is not a calculator. Jev does not reliably count, do math, or compare dates. Keep numbers and dates in regular software and give Jev only the judgment.

It reads questions literally. Vague questions get vague answers. Writing clear questions and answer options takes practice.

It only reads text for now. It does not handle images, audio, or video yet.

Extra information hurts it. Jev works best with only what it needs. Feeding it a customer's entire history to answer one small question makes it less accurate. Each request is also capped at roughly 64,000 tokens, less than frontier models allow.

"Cannot hallucinate" needs a footnote. Jev cannot return an answer outside the options you gave it. If your options are billing, technical, and sales, it will never invent "legal." It can still pick billing when the right answer was technical. It stays on the menu and can still order the wrong dish.

The benchmarks are mostly TypeSafe's own. TypeSafe used the average answers of other frontier models as its answer key, and no human checked those answers. On that evaluation, Jev landed around 68% agreement, roughly level with GPT-5.6 Terra and a few points behind Claude Opus 5.

TypeSafe says its headline "193x faster, 444x cheaper" figures are likely at the high end of real-world results. There is no public technical paper yet. Critics point out that fast classifiers that sort text into categories have existed for years. As far as anyone can tell, Jev tackles an old problem with a new architecture and a better developer product.

The Doom demo used a text description. Jev did not look at the screen. It received the game state as text, including enemy positions and distances. The demo shows fast decision-making. It says nothing about whether Jev can see a screen and play like a person.

Verification is on you. One YouTube review I watched summed it up well. Jev is incredibly fast and can even play games, but you end up verifying everything yourself, and that gets painful. I agree. Speed only pays off once you trust the answers, and that trust comes from testing against your own data.

It is early. Jev is in early access. Signups have been limited and were paused at one point because of demand. It runs only on TypeSafe's servers, so your data goes to them, which matters in regulated industries. TypeSafe has also said it cannot prove its current pricing is sustainable, although it expects prices to fall.

Where this goes

For the last few years, almost every AI product has been built around a conversation. You open a window, type a question, and read an answer.

Most of the work inside a business happens in systems and workflows, in thousands of small decisions nobody sees. Jev is one of the first models I have seen built for that layer.

It may not be the model that wins. OpenAI, Anthropic, Google, and the open model labs will likely ship decision models of their own, and community projects are already copying the approach.

The big labs are chasing general intelligence and consumer reach. TypeSafe is betting that making AI dependable inside ordinary business software is a big enough market on its own.

How I would start with it

I would not start by replacing anything.

I would pick one boring, high-volume decision that people make by hand today, like ticket routing or lead scoring. Then I would run Jev quietly alongside the existing process for a couple of weeks. People keep doing their jobs while the system records what Jev would have decided and how confident it was.

Then I would compare. Where did Jev agree with the team? Where did it disagree? Was its confidence lower on the cases it got wrong?

Only after that would I let it act, starting with high-confidence, low-risk decisions and keeping people in the loop for the rest.