OpenAI puts a visual, feedback-driven shopping interface inside ChatGPT

What OpenAI actually shipped

On November 24, 2025, OpenAI began rolling out shopping research to logged-in ChatGPT users on Free, Go, Plus, and Pro plans, across mobile and web. It is positioned for deeper product decisions—comparisons, constraints, and tradeoffs—rather than quick lookups like checking a price.

The flow is specific. You describe what you want, ChatGPT asks clarifying questions about budget, recipient, or features, then researches across the web and returns a buyer's guide with top products, key differences, and up-to-date details. The company says it takes a few minutes and performs best in detail-heavy categories: electronics, beauty, home and garden, kitchen and appliances, and sports and outdoor.

To seed adoption over the holidays, OpenAI is making nearly unlimited usage available to all plans through the season—a usage decision that reveals this is meant to become a habitual entry point, not an occasional novelty.

The interface is the product, not just the model

The notable engineering choice here is on the front end. Instead of a single prompt-and-answer exchange, OpenAI opens a visual interface where the research updates as it runs. Products come back mid-process, and users mark items 'Not interested' or 'More like this' to steer the results in real time.

In the background it looks across the internet for up-to-date information like price, availability, reviews, specs, and images, and brings options back to you as it goes.Montana Labs

This is a meaningfully different interaction pattern from chat. It treats research as a stateful session that incorporates new constraints as they arrive, rather than requiring the user to re-specify everything in a follow-up message. For applied teams, it's a working example of interleaving tool-driven retrieval with live user feedback inside a single task.

A small model post-trained for one job

Shopping research is powered by a version of GPT-5 mini—specifically GPT-5-Thinking-mini—post-trained with reinforcement learning for shopping tasks. OpenAI says it trained the model to read trusted sites, cite reliable sources, and synthesize across many pages, and built a new evaluation of difficult, constraint-heavy product queries.

Their accuracy metric is concrete: the percentage of products in a response that actually meet the stated requirements—price, color, material, specs. That framing matters because it measures whether the returned items satisfy constraints, not whether the prose sounds authoritative. OpenAI is candid that the model still makes mistakes on details like price and availability and directs users to merchant sites to confirm.

The commercial plumbing underneath

Two details signal where this is headed. Today users click through to retailer sites to buy; in the future, OpenAI says purchases will happen directly in ChatGPT for merchants in Instant Checkout. And merchants who want to appear in results can follow an allowlisting process.

OpenAI also states results are organic, based on publicly available retail pages, with chats never shared with retailers and low-quality or spammy sites avoided. Shopping research is additionally being surfaced proactively through ChatGPT Pulse for Pro users, which can suggest buyer's guides based on past conversations—for instance, e-bike accessories after e-bike discussions.

What a feedback-native shopping surface implies

The specific implication of this launch is that OpenAI is building a retail discovery layer that sits between shoppers and merchants, combining a purpose-trained small model, a real-time feedback UI, personalization from memory, and an emerging path to checkout and merchant allowlisting.

For teams building product-discovery experiences, the lesson is not that a frontier model can shop, but that OpenAI chose a mini model tuned for the task, wrapped it in a stateful visual interface, and defined success as constraint-satisfaction accuracy. The differentiation is in the interaction design and the retailer relationships, not raw model scale.