ChatGPT lets you try on clothes with a photo
Favorites keeps products and previews together, but the generated images don’t guarantee a garment will fit.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:003:49
Read transcript
ChatGPT is adding a virtual fitting room to shopping: shoppers can preview clothes and accessories on a photo of themselves, then save products and previews for later. OpenAI began the global rollout of Virtual Try On and Favorites on October first. A Try On button appears in shopping results, but you can also upload a product image— even a web screenshot—and ask for a preview using your own selfie or full-body photo. Favorites keeps the discovered item and its generated image together in an in-app Library. And shopping can start with an idea rather than a product: describe a style, or upload a celebrity outfit photo, and ask ChatGPT to find pieces to buy. The features use ChatGPT Images 2.5, which OpenAI released in September. The company says it better preserves reference-photo subjects and improves lighting, texture, and consistency when editing. OpenAI also claims up to fifty percent lower image-generation latency than Images 2.0. That’s a model comparison, not a promised wait time for a try-on. Most importantly, this is an appearance preview, not a guarantee that a garment will fit. TechCrunch reports OpenAI previously moved away from instant checkout after it performed poorly. These tools instead help shoppers evaluate and keep a product under consideration. Google shipped a similar try-on feature in July 2025, so the category is established. Whether people will use ChatGPT for it—and trust an AI-generated version of themselves when buying—remains an open question. That gap between what something looks like and what it actually contains showed up in a more literal way at autonomous-trucking company PlusAI. Two trailers stolen overnight from its Fremont warehouse held forty thousand pounds of sand, used as ballast to simulate real freight during testing—not Nvidia chips. Police recovered the trailers later that day. The powered tractor units carrying PlusAI’s driving technology remained inside the warehouse. Fortune reported that authorities said the thieves opened the trailers, found nothing valuable, and abandoned them. At least one trailer had Nvidia branding, but the motive is unknown. No arrests had been reported as of September 27. And in AI workflows, the measure that matters is often less the headline speed than how much waiting disappears. OpenAI’s GPT-6 Astra Ultrafast is available through its API and to eligible ChatGPT Work and Codex users; Amazon Bedrock offers another route. Nvidia says OpenAI used internal models to optimize the software serving Astra on Nvidia GPUs. Its claim of up to eight times faster token generation compares Ultrafast with Astra Standard. It does not mean an entire coding task finishes eight times faster. AWS cites a separate claim: up to six times faster API inference, with up to three hundred tokens per second. Those figures aren’t one universal guarantee across services. The practical test is whether faster output meaningfully shortens repeated code, tool, and review cycles. That focus on what happens after a first attempt also runs through new research on training agents to recover from mistakes. A method called PivotOPD has a teacher model show a student both the right action at a consequential wrong turn and recovery actions for the turns that follow. The researchers report the strongest average results against thirteen baselines across three benchmarks for two Qwen3 student models. They also report a 5.5 percent improvement over the strongest baseline on ALFWorld for one model, and a 3.2 percent increase in software tasks resolved on SWE-Bench Verified with a Nemotron student. Those are results on specific models and benchmarks, not a general guarantee. Across shopping, model speed, and agent training, the useful thing to watch is the same: whether a capability holds up in the messy steps between a promising first result and a decision people can rely on.
