jev·agent

Honest caveats

When not to use Jev

Jev is genuinely good at a narrow thing. This page is the other half — the limits, the failure modes, and the one claim that is weaker than it first reads. None of it is secret; most of it is just not the vendor's job to tell you.

Hard limits — these are not going to change with prompting

It writes nothing

No prose, no code, no summaries, no free-form anything. If your feature ends in text, you still need an LLM and now you are running two models.

Options are fixed per request

Choice picks from the set you send. It cannot propose an option you did not think of, so a genuinely open-ended question has no good encoding.

Text only

Text only — no image or audio input at launch Screenshots, documents as images and audio all need something in front.

~32,000 tokens per request per request

Long transcripts and big documents need chunking, windowing or retrieval first. An agent's full history will not fit.

Closed weights, hosted only

No self-hosting, no VPC, no air-gap. If your data cannot leave your network, Jev is out — see the open reimplementations instead.

Not OpenAI-compatible

No messages array, no choices[0].message.content. You cannot repoint an existing OpenAI SDK at it; you need a bespoke client.

The accuracy claim is weaker than it reads

This is the part worth slowing down on, because the headline numbers invite a conclusion the data does not support.

TypeSafe's published evaluation scores Jev at 67.8% — essentially tied with GPT-5.6 Terra (67.9%), and several points below GPT-5.6 Sol (74.1%) and Opus 5 (73.1%). So even the vendor is not claiming accuracy leadership; the claim is comparable accuracy at radically better cost and latency.

The strongest independent test so far ran 2,000 phishing emails against Claude Haiku 4.5 and found Jev's single verdict statistically worse (McNemar p < 0.0001), while winning decisively on speed and cost. The same study found the fix: five cheap signal questions in one call, combined in a plain logistic regression, reached 95.1% accuracy. Treat Jev as a signal generator you query broadly, not an oracle you query once.

Operational things that will bite you

Cases where it is simply the wrong tool

Open-ended extraction where you cannot enumerate the answers. Anything that must produce a human-readable explanation alongside the decision. Workloads under a strict data-residency rule. Tasks where a fine-tuned classifier on your own labelled data would win outright — if you have 50,000 labelled examples of exactly one task, a small trained model will likely beat a general decision model on both accuracy and cost.

Common questions

Can Jev replace my LLM?

No. Jev produces no text at all — no summaries, no code, no conversation. It replaces the decision steps inside a pipeline, not the writing. Most real systems run both.

Is Jev more accurate than a frontier LLM?

Not on the evidence available. TypeSafe's own table puts Jev at 67.8%, below GPT-5.6 Sol (74.1%) and Opus 5 (73.1%), and that figure measures agreement with other frontier models rather than correctness against ground truth. One independent head-to-head found Jev's raw verdict statistically worse than Claude Haiku 4.5.

Can I self-host Jev?

No. There are no open weights and no VPC or on-premises option. Community projects reimplement the interface with open models, but they are not Jev and their calibration is their own.

Can Jev handle images or audio?

Not at launch. Text input only.