Jev 1.14: Image Input, Box Localization, and Long Context

A support ticket can include a screenshot as well as a message. With the new image features, Jev can use both to answer a question such as “Is an error dialog visible?” With Jev 1.14, you can also locate a specific button or object in the image.

TypeSafe released the update on October 14, 2026. It combines model improvements with experimental features that developers enable separately. The distinction matters: Box requires 1.14, while image input and long context also work with 1.13. Source: release announcement.

The update at a glance

Change	Jev 1.13	Jev 1.14
Model improvements	Existing version	Broad improvements; TypeSafe expects the largest gain in multilingual support
Image input	Experimental, opt-in	Experimental, opt-in
box question type	Not supported	Experimental, opt-in
Extended context	Experimental, opt-in	Experimental, opt-in

TypeSafe calls the experimental features “danger zone” features. They are intended for exploring ideas and building prototypes; selecting the new model does not turn them on automatically. The release does not give a percentage improvement for multilingual tasks. Source: release notes.

Use images as context, instructions, or examples

Images can appear in three parts of a request:

• state: the screenshot or photo that questions refer to.
• instructions: a reference image relevant to a particular question.
• criteria: example images that help define the available choices or scoring levels.

The request lists images in assets and refers to them where needed. This lets you ask about a picture, compare it with a reference, or define categories visually. For image requests on either 1.13 or 1.14, add this header:

TypeSafe-Danger-Zone-Images: true

Images must be PNG or JPEG. Prepare files within the documented size limits before sending them; the API does not resize an oversized image to fit. Source: image documentation.

Possible applications include:

• Support triage: check whether a screenshot shows a particular error state.
• Product checks: compare a product photo with reference images.
• Visual classification: choose among categories illustrated with reference images.

Your application can use the answers directly to route tickets, flag items for inspection, or assign categories.

Locate a target with Box

A Choice question can select a category. A Box question locates one described target in an image, such as a button in a screenshot or a mug in a photograph.

The answer includes:

• box: the target’s top, bottom, left, and right edges.
• confidence: the model’s confidence in the location.

Coordinates are fractions from 0 to 1, measured from the top-left corner. To draw a box, multiply the horizontal values by the image width and the vertical values by its height. For example, left: 0.25 on a 1,200-pixel-wide image corresponds to 300 pixels from the left edge. Source: Box reference.

The JSON below defines a Box request. To send it as a complete multipart request, attach the image in a file field named screen-image and add the image header shown above:

{
  "model": "jev-1.14.0",
  "state": { "screenshot": "screen-asset-7c91" },
  "assets": {
    "screen-asset-7c91": { "type": "image", "part": "screen-image" }
  },
  "questions": {
    "button": {
      "type": "box",
      "instructions": "Locate the Submit button in `screenshot`."
    }
  }
}

Box does not take criteria. The question above returns one box around the described button, rather than a list of all buttons on the screen. Source: Box request format.

Context limits and pricing

The model documentation lists the same defaults for 1.13 and 1.14:

• Per request: 64k tokens for the state and all questions combined.
• Per question: 32k tokens for the state and one question.

Price shown on this site: $0.084 per million input tokens, with free output.

When a document exceeds the default limits, enable long context with a separate header. This raises both limits to 262k tokens:

TypeSafe-Danger-Zone-Long-Context: true

Both versions support this mode at the same per-token rate, though a longer input still costs more overall. TypeSafe also notes that longer inputs take more time and reduce accuracy. Enable it when the material you need exceeds the default limits. Source: long-context settings.

Images are billed as input tokens at the same rate as text, rather than at a fixed price per image. The response’s usage.input_tokens includes the image tokens. Source: image billing.

Which version should you select?

As of October 16, the documented model names are:

• jev-1.14.0: selects the new version explicitly.
• jev-preview: currently points to 1.14.
• jev-latest: currently points to 1.13.

The release announcement says jev-latest is scheduled to switch to 1.14 a week after release. To use 1.14 now and avoid changes when the alias moves, select jev-1.14.0 explicitly.

A straightforward approach is:

1. Keep existing text workflows on their selected version while you evaluate the update.
2. Select jev-1.14.0 when you need Box or want to explore the model improvements.
3. Enable Images or Long Context only for requests that use those features.

Jev 1.14 is also available through Vercel AI Gateway, where its model identifier is typesafe-ai/jev-1.14.0. Provider names and request formats can differ, so use the instructions for the channel you choose. Source: Vercel’s October 15 announcement.

Next steps

For an introduction to typed decisions, read what System One models and Jev do. For another model in the same category, see our Microsoft-Decision-1 guide.

You can also try Jev 1.13 on defapi. This is the 1.13 entry. For Box and other 1.14 features, choose a provider that supports the required version and feature flags.