Jev AI pricing explained: $0.042 per million, free output

What Jev charges, where AI providers spend their money, and why System One models can cost less.

Pricing checked 25 September 2026.

TL;DR: TypeSafe AI lists Jev at $0.042 per million input tokens, the small pieces of text it reads. Output tokens are free. Jev cost for a 3,000-token request is $0.000126. Run that same request 100,000 times and the bill is $12.60, before other services. Jev’s design can reduce running costs by avoiding step-by-step text generation and sharing input across questions.

I asked Jev to check a 2,000-word article against four sets of scoring rules. The whole request cost about $0.000126.

Jev is an AI model from TypeSafe AI. It makes decisions inside an app, such as choosing a category or giving a score. I used it to build a newsletter scoring dashboard. My hands-on Jev article shows the build and what it cost.

Use the tables below to estimate your bill. Then follow the money behind it: what providers pay for and which work Jev can avoid.

Jump to: Token costs · Provider costs · LLM output costs · Free output · LLM comparison · Real-time costs · Suitable tasks · FAQ

What Does Jev Cost per Token?

A token can be a word or part of a word. Jev charges for input tokens: the text you send, including your questions and instructions. Its output is the answer it sends back.

These prices use TypeSafe’s published Jev rate. All prices in this article are in US dollars.

That is 4.2 cents per million tokens, or $42 per billion.

To calculate a bill, divide the input token count by one million. Then multiply by $0.042:

Jev cost = input tokens ÷ 1,000,000 × $0.042

Count everything you send to Jev, including the text to check and the rules for checking it. Jev reports the input token count with each answer. Use that number to check the cost. A word count only gives a rough estimate.

Jev cost at different request volumes

A request is one job you send to Jev. It can contain several questions about the same text.

This table shows what you would pay at different levels of use. Each row assumes the same number of input tokens for every request, with no repeat attempts.

For a monthly estimate, use the number of requests you expect to make that month.

These figures cover Jev’s work. Add the cost of running your app, using other AI models, and paying any platform fees. If you send a request again, you pay for the input again.

What my four-question request cost

In my published Jev test, I sent a draft of about 2,000 words and asked Jev to score it on four things:

• How likely an AI search tool would be to use the article as a source.

• How easy it was to pull clear claims from the text.

• Whether it included evidence from my own work.

• How well the article was structured.

The reported input was 2,994 tokens.

2,994 ÷ 1,000,000 × $0.042 = $0.000125748

Rounded to six decimal places, that is $0.000126 for the whole request.

The request returned four judgments. Divide the cost by four and each judgment cost about $0.0000314. Sending four separate requests could cost a different amount because Jev may need to read the article each time.

The $5 credit I received would cover roughly 39,000 requests of this size.

This test shows what I paid for those scores. It does not show that the scores were accurate or that a higher score would bring more readers. My AI search checklist explains what I measure beyond a content score.

How Are AI Providers Charged?

An AI company can rent computers, buy its own machines, or pay another company to run a model for it. Each choice comes with a different bill.

Think of a print shop. You pay per page. The shop pays for equipment, electricity, and staff. AI has the same split between the unit on your bill and the costs behind it.

Your per-token price does not tell you which arrangement the provider uses or how much profit it makes.

Where the money goes

There are two stages to pay for. Training means teaching a model from data. Inference means running the trained model to answer a new request.

Training can involve preparing data, renting computers, testing results, and paying researchers and engineers. Providers may train updated versions later. AWS lists data preparation, training, hosting, and maintenance as separate costs.

Once a model is available, answering requests still needs resources:

NVIDIA explains the memory and processing demands and the power and cooling requirements. AWS’s cost guidance covers surrounding expenses. Some of these costs are included in a cloud rental rate, so they should not all be added again.

Why doing more work per hour lowers the cost

For an illustrative example, suppose a rented server costs $10 for one hour:

• If it completes 10,000 checks, the server cost is $0.001 per check.
• If it completes 100,000 checks in that hour, the server cost is $0.0001 per check.

Those are invented figures to show the calculation, not TypeSafe’s costs. They leave out the rest of the business.

The provider pays for the same hour but divides it across more work. NVIDIA uses this relationship between server costs and completed requests when estimating inference costs. Spare server capacity and quiet periods can push the average cost back up.

Why Do LLM Output Tokens Cost More?

Generating output tokens generally takes more time than processing input tokens. That helps explain why providers often charge more for output.

An LLM, short for large language model, is the type of AI used to write chatbot answers.

A typical text-generating model does two jobs. First, it processes your prompt. It can do much of that work at once because the input is already available.

Then it builds the reply in steps. Each new piece depends on what came before. The computers keep doing calculations and moving data from memory while the answer grows. NVIDIA describes these as the input-processing and text-generation stages.

A longer reply can therefore keep expensive equipment occupied for longer. The cost includes the work needed to produce the text.

Providers already reduce this work by reusing earlier calculations and handling several requests together. Those techniques help LLMs too. Their costs still depend on the model, the length of the request and answer, and how the service runs it.

Why Are Jev’s Output Tokens Free?

TypeSafe says Jev’s outputs are “too cheap to meter”, so it charges only for input. The answer still takes work to produce; free output is the company’s pricing choice.

Jev gives answers in a format you choose in advance. TypeSafe offers three answer types:

• Choice: pick from a list of options.
• Score: give a score using rules you provide.
• Noul: estimate how likely a yes-or-no answer is to be “yes.”

Why System One models can reduce running costs

System One models are built to return decisions that software can use. Jev is TypeSafe’s first model in this category.

TypeSafe says Jev works out the likelihood of its possible answers together. It skips the repeated text-generation steps described above. That is a change in how the model produces an answer, beyond simply asking a chatbot to write less.

It can also share work across questions. Jev reads the supplied information once and checks several questions against it in parallel. For my article-scoring task, that means sending one article with four questions in the same request.

The cost calculation above explains the benefit: if the same hardware completes more useful decisions in an hour, the cost per decision falls. Jev’s design provides a reason this can happen. It does not tell us TypeSafe’s exact savings.

TypeSafe also trains Jev specifically for decisions and probability estimates. That tells us what the training aims to achieve. It does not establish that training Jev was cheap.

The company’s own cost of running Jev is still unknown. Its price alone cannot prove that the service makes a profit.

Free output does not include writing articles or customer replies. Jev is built for decisions and scores. If your app needs written text, TypeSafe recommends using a model that can generate it. Include that model’s charges in your budget.

How Does Jev Pricing Compare With LLMs?

I would compare what each model costs to complete the same job.

A model with separate input and output prices charges for both the text you send and the answer it writes. Work out those two costs, then add them.

If both prices are listed per million tokens, the formula is:

Model cost = (input tokens × input price + output tokens × output price) ÷ 1,000,000

Keep the task the same. Choosing a category, giving a likelihood for every possible category, and writing an explanation take different amounts of work. Give each model the same information and decide what counts as a useful answer before comparing bills.

TypeSafe reports that Jev was about 194 times faster and 445 times cheaper in its own tests. Its explanation of those tests says the savings may be near the high end. It also notes that asking the other models to give probabilities for each possible answer adds work.

I would test my own task before using those savings in a budget.

The backup AI can cost more

Your app may send difficult cases to a second AI model. This is often called a fallback.

Suppose Jev handles 100,000 requests in a month. You send 5% of them, or 5,000 requests, to another model for extra help. For this example, assume that second model costs one cent per request.

The one-cent price is an example, not a quote for a specific model. This budget also leaves out app running costs and human review.

In this example, the second model costs nearly four times as much as Jev, even though it handles far fewer requests.

I would also calculate cost per usable result:

Total spent on the task ÷ number of results that meet your requirements

Include repeat attempts, the second model, and the time people spend checking or fixing answers. For customer support, a usable result might mean a message reached the right team without someone having to move it. For my dashboard, I would first need to check whether its scores help me edit better.

Is Jev Affordable for Real-Time Apps?

The number of requests matters as much as the size of each one. A small charge can add up when an app runs all day.

Suppose an app sends ten requests every second. Each contains 1,000 input tokens. At the listed price:

• One hour means 36,000 requests.

• Those requests cost $1.512, or about $1.51.

• Running at that rate all day and night for 30 days costs $1,088.64.

This is a calculated example. I have not tested an app running at that rate. The estimate assumes Jev can keep up and excludes other services.

TypeSafe reports response times of 70–500 milliseconds. That is 0.07 to 0.5 seconds. These are the company’s figures.

Test the response times in your own app, including the slowest replies. A screen that sorts support messages can tolerate a different delay from a system that must react at once. Jev’s response is only one part of the total wait.

Which Tasks Fit Jev’s Pricing?

I started by scoring articles because I could write down the questions I wanted Jev to answer. Choose one clear decision before choosing the model.

These are possible uses. Each still needs work around Jev’s answer.

Jev reads text. It does not directly accept images, audio, or video. To check a scanned invoice, another tool must first read the words in the image.

Use ordinary software for exact calculations. TypeSafe lists problems with counting, dates, distracting text, and instructions designed to mislead Jev. An answer can arrive in the correct format and still be wrong.

Start with one task you repeat often. Save the rules you gave Jev and which model version you used. Try examples from your normal work, and count the mistakes when you calculate savings.

Jev Pricing FAQ

Is there a free Jev plan?

I received a $5 sign-up credit for the test in my Substack post. That was the offer I received at the time. Check your TypeSafe account for the terms available to you.

A credit pays for usage until it runs out. It does not remove the usage charge. Free output also leaves you paying for the text Jev reads.

Does Jev charge separately for every question?

TypeSafe lists a price for input tokens, with no separate price per question. You can ask several questions in one request, as I did when scoring an article.

Extra questions add text, so they can increase the input charge. Use the token count Jev reports to check the bill. When several questions use the same document, sending them together can avoid paying to send that document repeatedly.

Can several questions share the same input?

Yes. TypeSafe lets you send one set of text with several questions. For example, Jev can judge a message’s category and urgency from the same message.

The questions must be answerable on their own. If the second question needs the first answer, wait for that answer before sending the next question. Asking independent questions together can save separate trips to Jev.

Can confidence scores eliminate review costs?

A confidence score shows how sure Jev is about an answer. I would test how well that score matches correct answers before removing human review.

TypeSafe recommends testing on your own examples. You choose a cutoff: answers above it can pass through, while answers below it need another check. The right cutoff depends on what happens when Jev is wrong.

Record how many answers pass without review and how many of those answers are wrong. Changing the cutoff changes both numbers. Saving on checks may create more work fixing mistakes later.

Is Jev’s low price sustainable?

We know what TypeSafe charges today. We do not know whether that price covers its costs and leaves a profit. The company says it needs time to show that the price can last.

I would keep the price easy to change in my cost estimate. I would also check what a higher rate would mean. That helps me plan; it does not mean I expect a rise. First, test whether Jev gives useful answers at the current price.

For the build behind these Jev cost calculations, read my hands-on Jev guide. Compare the token count from your own task with mine.

Subscribe to Product with Attitude for more free guides to AI costs and deciding what to automate. If you want to support more of this work, choose a paid Product with Attitude subscription.

All articles · Subscribe