TypeSafe AI Adds Vision and 256K Context to Jev 1.14

TypeSafe AI has updated its System One decision model with the release of Jev 1.14.0, expanding its single API endpoint at POST /v1/systemone. The update introduces three opt-in experimental capabilities under a danger zone program: spatial outputs, image processing within state and instructions, and a quadrupled context window. The underlying pricing structure remains unchanged at $42 per billion input tokens, equivalent to $0.042 per million, while output tokens remain free. The jev-preview alias currently points to version 1.14.0, with jev-latest scheduled to transition one week after launch.

Visual processing is enabled via the TypeSafe-Danger-Zone-Images header across both Jev 1.13.0 and Jev 1.14.0. Requests accept PNG and JPEG files up to a 32 MiB combined limit—including base64 encoding and JSON payload overhead—while rejecting animated PNGs. Users define visual assets once in a top-level map, referencing placeholders inside instructions, criteria, or state objects. For spatial tasks, Jev 1.14 introduces a Box output type where developers pass prompt instructions like the coffee mug in photo. The model responds with normalized top, bottom, left, and right spatial fractions ranging from 0 to 1 alongside a confidence score.

Because the Box feature emits four positional coordinates even if an object is missing, developers must adapt their implementation patterns. TypeSafe suggests executing a separate Noul boolean check to confirm an item exists before consuming the coordinate output. In internal evaluations focused on cutoff-detection across 36 cropped paintings, this multi-step approach proved necessary for edge-sensitive visual judgments.

For massive inputs, sending the TypeSafe-Danger-Zone-Long-Context header expands the standard 64,000-token request maximum and the 32,000-token state window by four times, raising limits to 256,000 and 128,000 tokens respectively. TypeSafe notes that while Jev 1.14.0 brings general performance boosts—notably in multilingual handling—reliability decreases as context expands. Consequently, practitioners moving beyond default windows should perform domain-specific benchmarks rather than relying solely on the higher token ceiling.