Jev Pricing: What It Costs Per Decision and Where It Is Free (2026)

Jev costs $0.042 per million input tokens, and output is free, so a typical short decision costs a small fraction of a cent. In our own test, 1,000 eight-way routing decisions on public data cost $0.0151 on Jev against $0.6089 on GPT-5.6 Terra. This page shows how the price is built, what we measured, where Jev is free right now, and how to estimate your own bill.

What TypeSafe charges for Jev

TypeSafe's Models page lists input pricing at $0.042 per million tokens, which is $42 per billion, with no separate charge for output tokens. There is no output to bill because Jev returns typed answers and probabilities, not generated text.

Two other routes exist. Vercel Developers announced on September 19, 2026 that Jev is free on its AI Gateway until September 25, so anything you run there before that date costs nothing at the model level. We ran our test through OpenRouter's decisions endpoint, which bills through OpenRouter. TypeSafe's own API has a waitlist. Prices and free windows change, so check the provider's page before you commit to a budget.

What you are actually billed for

The bill is input tokens only, and input is everything you send: the state, plus the text of every question, including its instructions and criteria. That has two consequences.

First, questions count. A request with one question and a request with five questions over the same state differ by the tokens in the four extra questions, not by five separate requests. Batching is the cheap way to use Jev, because the state is read once.

Second, long questions cost more than short ones. If you put a full policy into the criteria of a question, you pay for it on every call. Keep criteria as short as the meaning allows.

There is also a size limit. Per the TypeSafe docs, the state plus all questions in one request has a budget of 64k tokens, and the state plus your longest single question has a 32k budget.

What we measured

We ran the same 791 labeled decisions through Jev and four large language models on September 19, 2026, through one client and one billing meter, and recorded the cost each provider billed. This is a public-data test, not a client workload, and AY Automate has not used Jev on client projects. Cost per 1,000 decisions:

Against the two cheapest small models Jev was 4.7 to 7.5 times cheaper per decision, and against GPT-5.6 Terra it was 40 to 49 times cheaper. Those ratios are smaller than the headline multiples in launch coverage, because our test used short prompts and set the large models to minimal reasoning. The raw results are in the public benchmark files.

One detail explains why the gap is smaller than the price list suggests. Jev counted more input tokens than GPT-5.6 Terra for the same prompt: 360 against 153 on the 8-way task, and 952 against 828 on the 77-way one. Compare the billed cost per decision, not the price per million tokens.

A worked estimate

Say you route one million support messages a month, and each request is about 360 input tokens, as in our 8-way task. That is 360 million input tokens. At $0.042 per million, the monthly cost is $15.12. Our measured figure agrees: $0.0151 per 1,000 decisions is $15.10 per million.

Now double the questions per request. If the extra questions add 200 tokens per call, the same volume is 560 million tokens, or $23.52. Your bill grows with the words you send, not with the number of answers you get back.

Where the cost can hide

The model is cheap, so the surrounding system is usually where money goes. Retries add cost, so cap them. Fallbacks to a larger model add cost only for the items you escalate, which is why a confidence gate matters. In our test, using Jev alone when it was at least 80 percent confident and escalating the rest to GPT-5.6 Terra matched Terra's accuracy at 26 to 28 percent of Terra's cost. That is the pattern to price out: a cheap first pass, then a larger model only for the uncertain slice.

Engineering time is the other cost. A decision model needs labeled examples to test against, a threshold chosen on your own data, and monitoring for drift. Budget for that before you compare per-token prices.

When the price is not the point

If your workload is free-form writing, long reasoning, images or long documents, Jev is not a cheaper way to do it. It does not write text, it does not read images, and its context is limited. The price advantage applies to bounded decisions: routing, triage, scoring, guardrails, reranking and similar jobs.

For the launch framing of the price, see what Jev is and what the pricing means. For examples of what people have built at these prices, see our index of Jev use cases and examples, and for the patterns behind them, the guide to Jev use cases.

Keep going

• Measured accuracy, speed and cost: our independent Jev vs GPT and Claude benchmark.
• Build it: How to use Jev: a tested step-by-step guide with working code.
• What people built: the index of 1,300+ Jev builds.

FAQ

How much does Jev cost?

TypeSafe lists $0.042 per million input tokens with free output, which is $42 per billion input tokens. In our test, 1,000 short routing decisions cost about 1.5 cents.

Is Jev free?

Vercel Developers announced free access to Jev on its AI Gateway until September 25, 2026. Outside that window, and on TypeSafe's own API, it is billed by input token. Check the provider's page for the current terms.

Why is output free?

Jev returns typed answers with probabilities and does not generate text, so there are no output tokens to bill. You still pay for the input, which includes your questions.

How do I estimate my monthly Jev bill?

Multiply your monthly request count by the average input tokens per request, divide by one million, and multiply by $0.042. Measure the token count from a few real requests, because the questions count, not just the state.

Is Jev cheaper than GPT or Claude?

For short bounded decisions, yes. In our test it was 4.7 to 7.5 times cheaper than the two cheapest small models and 40 to 49 times cheaper than GPT-5.6 Terra per decision. It is not a replacement for a large model on writing or reasoning.

Does asking more questions cost more?

Yes, by the tokens in the extra questions. It does not cost a full extra request, because the state is read once and all questions are answered together.

What happens when the free window ends?

You pay the per-token price through whichever provider you use. If you built on the Vercel free window, price your workload at $0.042 per million input tokens before September 25 so the change is not a surprise.

How can I keep Jev costs low?

Batch questions over the same state, keep question criteria short, cap retries, and use a confidence gate so only uncertain items go to a larger model.