TypeSafe Jev 1.14 Ships Box Type for Object Localization and Opens Experimental Image Inputs

TypeSafe on Tuesday released jev-1.14.0, adding a fourth output type to its System One Model that returns structured bounding-box coordinates directly — no text generation, no downstream parsing. The Box type delivers {top, bottom, left, right, confidence} as floating-point fractions of image dimensions, making object localization a first-class output alongside Choice, Score, and Noul. The release also formally opens image inputs through an experimental "danger zone" opt-in, exposing multimodal capability present in the model architecture but deliberately withheld from the public API until now.

The combination extends the System One Model category — fast, structured decision-making borrowing from Kahneman's "thinking, fast and slow" framing — from classification into spatial reasoning for the first time. TypeSafe announced the release in a post published October 14, 2026.

Read more: TypeSafe's Jev delivers structured AI decisions without generating a single word

What the Box Type Returns and What It Costs

A developer declares a question with "type": "box" and receives coordinates directly in the API response:

{
  "model": "jev-1.14.0",
  "answers": {
    "mug": {
      "type": "box",
      "box": {"top": 0.25, "bottom": 0.82, "left": 0.44, "right": 0.84},
      "confidence": 0.92
    }
  },
  "usage": {"input_tokens": 290, "output_tokens": 21}
}

Coordinate values are normalized to image dimensions — 0.44 left means 44% from the left edge — making the output resolution-independent. Any downstream system multiplies the fraction by the actual pixel dimension. The confidence score reflects TypeSafe's RLCD (Reinforcement Learning for Calibrated Decisions) training, which the company says calibrates confidence to actual accuracy rather than raw model logits, as described in TypeSafe's model documentation.

Output cost is negligible: the example above uses 21 output tokens. Under TypeSafe's pricing ($0.042 per million input tokens; output tokens free), localization adds almost no marginal cost over classification.

Images Enter the API as Composable Assets — With a Warning

TypeSafe's image-input system treats images as composable elements that can appear anywhere in the request schema — in state, instructions, or criteria. Images are referenced in the JSON body via UUID placeholders and supplied as multipart form-data, as detailed in TypeSafe's image composability documentation.

The TypeSafe-Danger-Zone-Images: true header is not a formality. TypeSafe explicitly describes danger zone features as capabilities that do not yet meet its production quality bar and warns that multimodal performance is "more variable" than text-only inference. The same release ships a TypeSafe-Danger-Zone-Long-Context header that bypasses the production context ceiling. Both are opt-in because TypeSafe is shipping faster than its normal stability guarantees allow. Supported formats are static images — PNG and JPEG only; video and animation are not supported.

The Model Was Already Multimodal

TypeSafe acknowledges that the architecture underlying jev-1.13 already accepted image inputs at the model level. The 1.14 release is not about adding multimodal architecture — it was already there — but about exposing it under experimental conditions. The company held back a working capability for at least one model version, presumably because performance characteristics were not yet stable enough for public release. Making it a danger zone feature is a middle path: available to developers who need it, but without the full quality guarantee of the core text-only decision types.

Spatial Reasoning in a Category OpenAI Now Competes In

TypeSafe created the typed decision API category but is no longer alone in it. OpenAI launched its Decisions API (GPT-6 Luna) in public beta on October 6, 2026 — announced at DevDay 2026 on September 29 — with image support included from the start, no experimental flag required. AWS has Strands Decider 2B; Cloudflare operates Clef and Clef-flash in adjacent territory.

Read more: vLLM-SR releases Decision 3.0, an open multimodal suite that tops Jev on text and vision

TypeSafe's danger zone shipping strategy appears partly motivated by that competition — exposing experimental capability so developers building vision workflows have the option before the full production bar is reached.

The meaningful differentiation TypeSafe holds is the Box type itself. OpenAI's Decisions API does not have a documented equivalent: image inputs return structured decisions, but not typed bounding-box coordinates as a first-class output. For workflows that need localization rather than classification, TypeSafe currently occupies territory no competitor has formally entered. The practical advantage over generalist vision-language models is the elimination of a parsing step: a Box response delivers coordinates consumable directly by any drawing, cropping, or detection system.

Availability and What to Watch Next

jev-1.14.0 is available now via the jev-preview model identifier. The jev-latest pointer — used by most production integrations — is expected to advance to 1.14.0 approximately one week later. Developers who need Box outputs before that window must explicitly specify jev-1.14.0. The Box type requires this version specifically; it cannot be accessed on earlier model versions regardless of header settings.

TypeSafe has not announced a timeline for moving image support or long-context out of experimental status. The sharper question for the category is whether competitors will add localization outputs to their own decision APIs. The Box type TypeSafe shipped today is a natural extension of any decision API that accepts images — which means every competitor now has a concrete capability gap to close.