TypeSafe AI Releases Jev 1.14 With Images, Box Type and 262k Context

TypeSafe AI released Jev 1.14, a System One model that returns typed decisions — choices, scores, yes/no-style answers with probabilities and confidence — instead of free text. The new version, live under the jev-preview alias, adds three experimental features behind opt-in "danger zone" headers: image input, a Box question type that returns a normalized bounding box, and a long-context mode up to 262k tokens.

Pricing stays at $0.042 per million input tokens with free output. The jev-latest alias moves to 1.14 in a week, and 1.13 stays available with 60 days' notice before deprecation.

We appreciate the candor. A vendor labeling its own features "danger zone," warning that long context is less reliable, and telling developers to trust their own evals over its internal charts is a refreshing change from benchmark-maximalist launches.

The typed-decision approach is genuinely useful for automation. If a workflow needs a classification with a confidence score, getting structured output by design beats parsing prose and hoping.

The Box type shows the risk of moving into vision: it always returns a box, even when the object isn't there. TypeSafe recommends pairing it with an existence check, and its small test shows edge checks still missed some cut-off crops. In inspection settings, a confident box around nothing is exactly the error that slips past a busy reviewer.

I'm watching whether teams heed the not-for-production label, or whether the cheap pricing pulls these features into real pipelines early.