What is Jev? How to use it in Sigma through MotherDuck
Let's start with a basic scenario where a contact center logs about 2 million notes a month. Analyzing the data is time-consuming, and most people can read only a few hundred before they run out of time.
One solution might be to run an LLM over all the contact center logs, but at that volume the costs start to add up. In fact, if you have 10 questions you want to ask of the data, you're looking at a bill around $10,000 a month. Using Jev, you can run those same 10 questions for a fraction of the cost, about $100.
Sigma's AI columns, which are in public beta, let a business user ask a question in plain language and get an answer on every row of a workbook. To see how Jev performs in that role, we run it from a Sigma workbook on MotherDuck. Jev tags 6,578 customer reviews with the key product features each one references, in 4.6 seconds for 5.3 cents.
What is Jev?
Jev is an AI model from TypeSafe AI built for text classification. It answers closed questions about text and writes no text of its own. When you give it a record and a question with a fixed set of possible answers, it returns one answer, with a probability for every option. TypeSafe says a call takes about a tenth of a second and costs a sliver of what a general-purpose LLM charges.
The way that Jev works is that it reads the whole question in one pass and returns a structured answer. That answer is a yes-or-no probability, a pick from a list you write, or a spot on a scale you define. Because the answer is structured, software can act on it without parsing a sentence.
Jev is trained to give calibrated odds: an answer at 90% confidence should be right about nine times out of 10. The amazing thing is the price. TypeSafe lists it at $0.042 per million input tokens, with output free.
Because it doesn't generate text, Jev can't summarize a call or explain a choice. It also sees nothing but the record you send, so a ticket that reads "same issue as last week" gives it nothing to work with.
Why Jev matters when it runs inside your data platform
Without Jev, the choices for labeling text are a custom classifier someone has to train and maintain, or an LLM whose price rules out running it on every row. Because MotherDuck exposes Jev as a SQL function, its label lands in the same query that reads the text. You can filter and group on that label like any other column.
Because the cost is a fraction of a cent, a team labels every comment from the quarter instead of a sample, and total AI spend rises. Finance will want each labeling workload tagged so the bill can be traced.
What the numbers show when Jev runs in Sigma on MotherDuck
The test runs from a Sigma workbook through the MotherDuck connection, with prompt_jev() tagging each of 6,578 customer reviews with the key product features it references. As mentioned earlier, it finishes in 4.6 seconds and costs 5.3 cents.
| Line item | Amount |
|---|---|
| Rows categorized | 6,578 |
| Estimated tokens | 842,033 |
| Run time | 4.6 seconds |
| Compute on a Standard MotherDuck warehouse ($0.0007 per second) | $0.003 |
| MotherDuck AI processing fee | $0.05 |
| Total | $0.053 |
Figure 1. What Jev's run on 6,578 rows costs.
The same 842,033 tokens would cost about $1.17 on Claude Haiku 4.5, Anthropic's lowest-priced current model, at its list rates. That estimate assumes about 10 output tokens per review. The contact center from the start of this post pays about $16 to ask Jev one question of 2 million contact notes, against about $356 on Haiku.
| Job | Jev cost | Jev time | Claude Haiku 4.5 cost (estimated) |
|---|---|---|---|
| Our test: 6,578 reviews, 1 question | 5.3 cents | 4.6 seconds | about $1.17 |
| 100,000 rows, 1 question | about 81 cents | about 70 seconds | about $18 |
| 2 million notes, 1 question | about $16 | about 23 minutes | about $356 |
| 2 million notes, 10 questions | about $160 | about 4 hours | about $3,560 |
| A year of those 10 questions | about $1,940 | about 47 hours | about $42,700 |
Figure 2. Jev figures scale our run linearly. Haiku figures price the same token count at Anthropic's list rates ($1 per million input tokens and $5 per million output tokens), assuming about 10 output tokens per row. The Haiku figures are estimates rather than a test run, and real costs depend on record length, prompt size, and model.
How you should be thinking about Jev (and how we are thinking about it)
Most teams should start with AI columns. Each idea becomes a column you can try and throw away in minutes, with no ticket to the data team. The prompt runs through the data platform's own AI function on Snowflake or Databricks, so the data stays put and the column keeps the access rules on the connection. Sigma caches every answer and reruns the model only when a row changes, and admins set a monthly token cap per connection. The default cap of 10 million tokens covers about 78,000 reviews the size of ours, or about a dozen runs of our test, before anyone has to ask about the bill.
Try Jev in Sigma
If your data lives in MotherDuck, ask your Sigma account team for access to the private beta of the MotherDuck connection. MotherDuck's guide to prompt_jev() walks through writing your first label list. Teams on Databricks or Snowflake can add an AI column with a fixed set of categories to the free-text column they read least. If you're not a Sigma customer, we'd be glad to show you how it works on your own data.
Frequently asked questions about Jev in Sigma
Is Jev a large language model?
Jev is a classification model rather than a large language model. It returns a structured answer, such as a pick from your list or a yes-or-no probability, with odds attached to every option.
How much does Jev cost?
TypeSafe lists Jev at $0.042 per million input tokens, and output is free. A token is three-quarters of a word on average, so labeling a long customer comment costs a tiny fraction of a cent.
Is Jev more accurate than an LLM for text classification?
In an independent phishing test, Jev trails Claude Haiku on one broad question, 63% to 81%. Its accuracy rises to 95% when the testers split that question into five narrow ones. Test Jev against labels you trust before you rely on it.
How do I use Jev in Sigma?
Connect Sigma to MotherDuck, which is in private beta, and call prompt_jev() from a workbook. Use a custom SQL query when you need a list of categories, or a formula column for a yes-or-no question. Teams on Databricks or Snowflake can call Jev through a passthrough function and a function in their data warehouse that wraps TypeSafe's API.
Does my text stay inside MotherDuck when I use prompt_jev()?
Your text goes to TypeSafe for processing, because MotherDuck calls the TypeSafe API on your behalf. MotherDuck admins can turn the function off. The integration is in preview, so clear it with your security team before using it on sensitive data.
