OpenAI previews Decisions API and adds virtual try-on to ChatGPT
The AMW Read
These incremental product additions extend a known foundation-model provider into specialized classification and shopping distribution, with segment-level implications for competing service providers.
OpenAI previews Decisions API and adds virtual try-on to ChatGPT
OpenAI introduced its Decisions API at DevDay, offering the Luna model in limited preview to select from predefined options, including image categories and agent behaviors. Sam Altman said restricting the output to a single choice enables fast responses while preserving image understanding, broad language support and safety protections. Separately, ChatGPT added global virtual try-on and Favorites features powered by ChatGPT Images 2.5. Users can upload photos to preview clothing and accessories, then save products and try-on images to a Library.
The releases extend OpenAI's position from general-purpose models into specialized decision services and consumer discovery. The Decisions API overlaps with TypeSafe AI's Jev classifier, putting a frontier lab in competition with a startup focused on fast, inexpensive classification. This creates a practical differentiation question: whether specialized providers can offer better decision quality per dollar. Shopping adds another distribution surface, competing with Pinterest and Google after OpenAI's reported retreat from instant checkout. The source also reports three safety researchers were dismissed over alleged confidential information sharing; that is a separate governance issue, not evidence of either product's safety performance.
For builders, the concrete implication is to evaluate constrained-choice models separately from general-purpose generation. A Jev hackathon demo checked agent actions against assigned tasks, illustrating a potential oversight use case, but it does not establish production reliability or prove that Luna offers equivalent controls. Teams considering the Decisions API should compare classification accuracy, latency and total operating cost on their own workloads before using it to approve or block agent actions.
