Jev and System One models: what they mean for sales teams
TypeSafe's Jev answers typed questions in milliseconds instead of writing text. What it is, where it fits in a sales stack, what it cannot do, and how to test it on your own pipeline.
On 15 September 2026, TypeSafe AI released Jev into early access, alongside a $40 million seed round led by DCVC. TypeSafe was founded by Diogo Almeida, a former OpenAI researcher, and Jev is the first model in a class the company calls System One models.
Most coverage has been about the speed claims. For a sales team the more useful point is simpler. Most of the AI in a sales stack is not writing anything. It is deciding: which rep gets this lead, is this reply a meeting request or a brush-off, is this deal stalling. Those are the jobs Jev was built for, and it is worth knowing where it fits before a vendor tells you.
What Jev actually is
A language model answers in prose, one token at a time, and your software has to read the prose back. Jev skips the prose. You send it a block of state, for example an email and the account’s record, plus a set of typed questions, and it returns typed answers:
- A choice from a list of options you define, such as which of six queues a lead belongs in.
- A score against a rubric, such as how urgent a reply is.
- A yes or no, such as whether a message asks for a meeting.
Each answer comes with a calibrated probability. Calibrated means the numbers are meant to be honest: across many decisions, the ones Jev gives 90% should be right about nine times in ten. That is what makes it usable in software. You can act automatically above one threshold and send everything below it to a person, instead of trusting every answer the same.
The name is a double reference. System One comes from Daniel Kahneman’s fast, intuitive System 1 thinking. Jev comes from William Stanley Jevons, whose paradox says that when something gets cheaper, people use far more of it.
The numbers, and how to read them
$42
| Claim | What TypeSafe reports |
|---|---|
| Latency | 70 to 500 milliseconds, end to end |
| Speed against a language model | 20 to 200 times faster on decision-shaped workflows |
| Cost against a language model | 40 to 400 times cheaper on the same workflows |
| Price | $42 per billion input tokens, output free |
The multipliers come from TypeSafe’s own comparisons on its chosen workflows. Read them as what is possible on a decision-shaped task, not as a promise for yours. The latency range is the number that matters for anything live: under half a second is fast enough to use while a buyer is still talking.
Where it fits in a sales stack
The test is whether the job is a decision with a known set of answers, made many times.
| Job | The typed question | Why it fits |
|---|---|---|
| Routing inbound leads | Which rep or queue should take this? | Runs on every form fill, and the options are fixed |
| Sorting replies | Meeting request, objection, referral, unsubscribe or out of office? | A choice from five, thousands of times a week |
| Deal risk | How stalled is this deal, given the last month of activity? | A score against a rubric you already use |
| CRM hygiene | Does this reply say the contact has left the company? | A yes or no, with a threshold before anything changes |
| Buying signals | How strongly do these notes point to a purchase this quarter? | A score you can sort a list by |
| The live call | Which kind of objection did the buyer just raise? | A choice that has to be made in well under a second |
What it cannot do
This is where most of the value of knowing about Jev is, because it keeps a team from buying the wrong thing.
- It does not write. No emails, no call replies, no summaries, no chat. Anything a buyer reads still needs a language model or a person.
- It does not reason in steps. A multi-part judgment has to be broken into single questions in your own code.
- It reads text only. No audio, images or video, so a call has to become a transcript or captions first.
- It does not explain itself. You get a number, not a reason, which matters when a manager asks why a lead was routed away from them.
- It only knows what you send it. It does not browse or look anything up, so the account context has to be in the request.
- It can struggle with arithmetic, counting and dates. Keep those calculations in your code.
Decide fast, then write
The useful pattern is two steps. A fast model decides what is happening and whether anything needs to happen. A generative model, or a person, writes only when it does. Most messages need a decision and never a paragraph, so the expensive step runs far less often.
A live sales call is the clearest case, because the time budget is a few seconds. Recognizing that the buyer just raised a timing objection is a decision, and a System One model can make it in a fraction of a second. Writing a reply that mentions the buyer’s own business, in the seller’s voice, is generation, and it is where the rest of the time goes. That is the split Axelize is built around: it recognizes the moment, then writes the sentence while the buyer is still talking.
On a discovery call, Johanna Lorenz from Fourth Coffee says: “We'd love to, but let's revisit this after the holidays.” Axelize finds: Fourth Coffee is opening four new stores in November, and its hiring page lists twelve open roles for them. It suggests: “Totally understand. You're opening four stores in November, though, so is the question whether the new teams start on this or get moved over later?” Recognizes a timing objection, then ties it to a date the buyer already has, so waiting has a visible cost.
How to test it on your own pipeline
- Pick one decision your team makes more than 100 times a week. Lead routing and reply sorting are the usual first choices.
- Write it as one typed question with the exact options you would accept, including an “other” or “none” option.
- Pull 50 to 100 past examples where you already know the right answer.
- Check the calibration. Of the answers it gave at around 90%, were about nine in ten right? If not, the probabilities cannot drive automation yet.
- Set two thresholds. Act automatically above one, send to a person below the other, and look at what lands between them.
- Cost it. Tokens per decision, times weekly volume, times $0.042 per million. Sorting 10,000 inbound emails of about 500 tokens each is 5 million tokens, which is about $0.21 at the listed price.
Sources
The figures above are TypeSafe’s own, as published in its announcement and reported in the coverage below, as of 26 September 2026: Introducing System One Models and Jev; the funding coverage in Forbes and SiliconANGLE; and DataCamp’s explainer, Jev: TypeSafe’s System One Model.