TypeSafe AI has launched Jev 1.14, a System One model designed to provide typed decisions rather than standard generated text. The model utilizes Reinforcement Learning for Calibrated Decisions to return structured answers, including probabilities and confidence scores for each query.
The update introduces support for image processing, allowing users to reference images within state, instructions, or criteria. Developers must enable the TypeSafe-Danger-Zone-Images header to utilize this feature. The company notes that model performance on images is currently lower than on text-based inputs.
A new Box question type provides bounding box coordinates as fractions ranging from 0 to 1, accompanied by a confidence score. TypeSafe recommends combining this with Noul questions to verify object existence, as the model will return coordinates even if the target is absent.
The release expands context limits to 262k tokens for both requests and individual questions when the TypeSafe-Danger-Zone-Long-Context header is active. Previously, these limits were capped at 64k and 32k tokens respectively to maintain reliability.
Pricing for the service remains at $0.042 per million input tokens, with output tokens provided at no additional cost. The company advises developers to rely on their own evaluation metrics, noting that internal performance charts may be biased toward their own models.
