Jev: Stop Paying a Chatbot to Answer Yes or No

Stop paying ChatGPT to say yes. A lot of apps send a full chatbot prompt just to get a yes, a no or a category back. Jev is a model built only for that job: it can't write a single word, it returns a decision with a confidence score, and its output is free.

What happened

Jev comes from TypeSafe AI, a San Francisco lab. It was released on September 15, 2026. On October 9, TechCrunch reported the company raised $870 million at a $7.5 billion valuation, led by Andreessen Horowitz, with Sequoia and existing investor DCVC participating.

InfoQ's write-up from October 1 explains how it works. "Jev does not generate text. It returns typed, probabilistic decisions that software can act on directly." You send it a state, which can be a string or structured data, plus a set of typed questions. It answers all of them in one parallel pass. Each answer comes back as a typed result, such as a choice or a score, with a probability distribution and a confidence value. Your code acts when the confidence is above a threshold and escalates to a human or a bigger model when it's below.

The numbers InfoQ reports:

• Input costs $0.042 per million tokens. Output is free.
• The context window is 32,000 tokens.
• TypeSafe quotes 70 to 500 milliseconds end to end. OpenChamber analyzed launch posts and found a median user-reported latency of 76 milliseconds.

Early uses InfoQ lists include safety classification at Vercel, email classification, model routing, and screening agent tool calls before they run. It integrates with Vercel AI Gateway, Netlify and LangChain.

The limits

TypeSafe documents its own weak spots: unreliable counting, arithmetic and date comparison, and accuracy loss on large, noisy inputs. Its advice is to keep math in code. A Hacker News commenter quoted by InfoQ put the main risk well: Jev "cannot emit an invalid type but can still emit a completely wrong valid value." InfoQ also recommends pinning a specific version, such as jev-1.13.0, and avoiding moving aliases like jev-latest.

Why it matters for builders

Look at the LLM calls in your own product. I'd bet a good share of them are classification wearing a chatbot costume: is this email spam, is this ticket urgent, does this message break the rules, which model should handle this request. You pay for output tokens on every one of those, you parse free text that sometimes comes back malformed, and you wait for generation.

A typed decision fixes all three. You can't get an unparseable answer back, you get a confidence number you can route on, and there's no output to pay for. The tradeoff is that you have to trust a calibrated probability, and you own the threshold.

Try it in 15 minutes

1. Search your codebase for prompts that end in a constrained answer. Phrases like "answer yes or no," "respond with one of," "classify this," or a JSON schema with a single enum field are good signs.
2. Pick the one that runs most often. Log 50 to 100 real inputs alongside the answer your current model gave.
3. Get an API key and open the Jev playground. Recreate the question as a typed choice with the same options.
4. Run your logged inputs through it and compare. Note where the two disagree and look at Jev's confidence on those cases.
5. Pick a threshold. Above it, act on Jev's answer. Below it, fall back to your current model. That hybrid is the safe first deployment.
6. Pin the version and keep your comparison set, so you can re-run it when Jev updates.

Here's how you get paid

My opinion: lots of apps are burning money on chatbot calls for simple yes-or-no decisions, and most of the teams behind them haven't looked.

That's a cost-swap service. Audit a client's LLM usage, find the classification calls, run the side-by-side comparison above on their real traffic, and swap in a typed model where it holds up, with the fallback in place. Then propose a fee based on a share of the savings for a set period. The client takes no risk, the comparison data makes the decision for them, and you get paid in proportion to what you save.

Keep the audit honest. Some calls really do need a full language model, and the report should say so.