news

OpenAI puts ChatGPT Images 2.5 to work on virtual clothing try-ons, with no accuracy figure

The October 1 release note adds a Try on button to clothing and accessory listings and stores your reference selfie, while saying nothing about how often the generated fit is right.

By Daniel Ochoa · 1 October 2026 · Updated 4 October 2026

OpenAI moved its image model into retail on October 1, 2026, adding a Try on button to clothing and accessory listings inside ChatGPT. Tap it, take or upload a selfie, and ChatGPT Images produces a picture of you wearing the item. The company announced a global rollout on web and mobile, paired with a Favorites function for saving products into the ChatGPT Library, and said both features run on ChatGPT Images 2.5.

What the release note commits to, and what it does not

The October 1 entry in ChatGPT's release notes is short. It states that a Try on button will appear on product listings for clothes and accessories, that selecting it prompts a selfie, and that the generated try-on comes from ChatGPT Images. It adds that the reference photo is saved for future try-ons and can be changed or deleted under Settings, Personalization, Reference photos. There is no eligibility detail in the note about which accounts, plans, countries or merchants get the button, and no numbers of any kind.

The model backing it shipped three weeks earlier

ChatGPT Images 2.5 arrived on September 8, 2026, promoted by OpenAI on sharper detail, more precise editing and faster generation, with latency cut by half and two API variants split along a speed and quality line. Try-on is the first consumer feature to be publicly pinned to that model as its reason for existing, which makes the shopping surface a live demonstration of the September release rather than a separate product bet.

A generation task that can actually be graded, shipped without a grade

Most image-generation quality claims are argued over aesthetics, which is why the category settles disputes with blind preference votes like those collected in public image arenas. Virtual try-on is different, and more interesting, because it has a checkable answer. The garment has a real cut, a real drape and a real colorway, the shopper has a real body, and the delivered package either resembles the generated preview or it does not. That is a measurable error rate waiting to be measured, by garment type, by body type, by lighting.

Nothing in the announcement supplies one. There is no stated fidelity rate, no comparison against the try-on features already shipped by other platforms, no breakdown by category, and no indication of how the output was evaluated before launch. For a feature whose entire value proposition is predictive accuracy about a purchase, the absence of a number is the story. Retailers measuring this already tend to report it through downstream proxies such as purchase intent and return rates, which are the figures that will eventually test whether the generated previews hold up.

The reference photo is persistent by default

The storage behavior deserves attention independent of image quality. OpenAI's note says the selfie is kept for future try-ons, with change and delete controls sitting several menu levels deep in Settings. That is a biometric-adjacent reference image retained by default on a consumer shopping surface, which is a materially different privacy posture from a one-off generation that discards its input. The controls exist, and the note tells you where they are, but the default is retention.

Sources: OpenAI, TechCrunch.