Hi, and welcome to the first issue of The Jev Weekly.

Jev has been public for four days, and a lot has happened. Here's the short version.

What changed

Jev launched. On September 15, TypeSafe AI released Jev, the first "System One" model. You send it state and typed questions (Choice, Score, or Noul) and it returns answers with calibrated probabilities instead of text. It costs $0.042 per million input tokens, and output is free. Access is through an early-access waitlist, and early users report getting in within a day or two. (Launch post · Our explainer)

$40M seed round. TypeSafe came out of stealth the same day with a $40 million seed led by DCVC. (Announcement)

Two SDK releases, both with breaking changes.

• JavaScript SDK 0.6.0: Score criteria are now an ordered array instead of an object keyed by integers. (Release)
• Python SDK 0.7.0: switched from msgspec to Pydantic, and system_one() gains a response_model argument that parses answers straight into your own Pydantic model. (Release)

TypeSafe published Jev's weak spots. A new docs page lists nine known failure modes of jev-1.13, including arithmetic, date comparison, multi-hop questions, and large, noisy state. It's worth reading before you build anything. (Jev 1.13 jaggedness)

Rate limits are 250,000 tokens per second and 1,200 requests per minute, and TypeSafe says they're adjusting dynamically while demand settles. Pin jev-1.13.0 instead of jev-latest once you've tuned anything.

This week in Jev projects

Our weekly search found more than 1,000 GitHub repositories mentioning Jev in its first four days. We've reviewed the notable ones and listed 71 on mrjev.com/projects. Five worth your time:

• jev-ultrafast (3.6k stars): Browser Use's fast browser agent. Jev picks every action; a small model only types text. The authors report a full Google Flights search in about 7 seconds.
• fast-jev-compaction (2k stars): a Claude Code plugin that replaces the lossy compaction summary with Jev deciding, call by call, what to keep verbatim.
• OpenJev (1.2k stars): can you run something like Jev on a single RTX 3090? A frozen 4B model says: sort of. Great for understanding how System One models work.
• pg-jev: a PostgreSQL extension that lets you filter and classify rows with plain-language conditions.
• TypeSafe Mario: Jev plays Super Mario Bros. from emulator RAM, no screenshots. The most fun demo of the week.

Using Jev from Claude Code or Codex? Start with the MCP servers.

One practical tip: don't trust "100% sure" until you've measured it

Jev returns a confidence score with every Choice and Score answer, and it's tempting to automate anything above 0.9. The Janus project checked what happens when Jev reports full confidence. Its author found those answers were right 95.8% of the time on the Banking77 intent dataset, but only 76.7% of the time on Web of Science paper abstracts.

Same model, same confidence, very different accuracy. Calibration depends on your data. Before you let confidence gate a real action:

1. Label 100 to 200 real examples.
2. Run them through Jev and plot accuracy by confidence band.
3. Pick the threshold that meets your error budget, and route everything below it to a person or a stronger model.

jevcal automates exactly this, and can fail your CI when a model update shifts the numbers.

Try the numbers yourself

Wondering what Jev would cost for your workload? The cost calculator compares it with Claude and GPT models at current list prices.

That's it for this week. Hit reply and tell me what you're building with Jev. I read every reply, and the most interesting ones may end up in a future issue.

MrJev

MrJev is independent and not affiliated with TypeSafe AI. Numbers attributed to projects are their authors' own.