Jev, from TypeSafe AI, doesn’t generate text. It’s a classifier: you give it an input and a set of JSON rules, and it returns decisions and scores, such as churn risk or whether an email needs a reply.
TypeSafe says Jev is 20 to 200 times faster and 40 to 400 times cheaper than typical models, with output tokens free. I tested that claim on 1,500 of my own exported emails. This video is not sponsored.
Summary
- Jev scored four fields for each email: category, priority, spam and whether it needs a reply.
- A 100-email batch on eight workers averaged 200 ms per email, with a p95 of 240 ms and about 38 emails per second.
- It rated account and security notices, such as a domain revocation, as high priority, and gave an account-violation email a 91% chance of needing a reply.
- Scaling to 1,000 emails kept the same roughly 200 ms average for the full round trip, across more than 1.1 million input tokens.
- The run cost about $0.05.
How to apply it
- Use a classifier like Jev when you need a structured decision rather than prose: triage, spam filtering, routing or risk scoring.
- Define a few fields that map directly to actions, such as “needs reply” and “spam,” instead of many vague labels.
- Start with a 100-item batch, check the highest-scoring results by hand, then scale up.
- Run requests in parallel. Per-item latency stays flat, so more workers means more throughput.
- Try it at typesafe.ai or read the launch post.
Verdict
Remarkably fast and cheap for structured decisions. It’s not a replacement for a text model.
What’s good
- About 200 ms per decision, even across 1,000 emails
- A full test run cost about $0.05
- Sensible priority and reply predictions on my inbox
What’s not
- Returns scores and labels only, with no text generation
Get the next test the day it’s out.
