ChatGPT Images Evolves From Editing Tool to Visual Ad Platform

From a faster editor to a core product

OpenAI announced "the new ChatGPT Images" in December 2025. At the time it looked like a routine model upgrade. Ten months later, it looks more like the start of OpenAI's strategy for visual media. The December release added a new flagship image model to ChatGPT for all users. OpenAI said the model made precise edits without disturbing the rest of an image and generated results up to four times faster. The same model went to developers as GPT Image 1.511. It also added a dedicated Images area in the ChatGPT sidebar, with preset filters and trending prompts meant to spark ideas without requiring a written prompt11.

The pitch to businesses was clear from the start. OpenAI said GPT Image 1.5 kept branded logos and key visuals more consistent across edits. It aimed the model at marketing graphics, logo work and ecommerce teams building full product catalogs (variants, scenes and angles) from a single source image11. It also cut API image input and output prices by 20% compared with GPT Image 111. Wix's head of AI research said the model supported "faster concept-to-production workflows" and called it one of the leading image models available11.

Since then the product has moved quickly. GPT Image 2 arrived in April 2026 and added reasoning to image generation12. Wired found that Images 2.0 could search the web, produce several images from one prompt and handle aspect ratios from 3:1 to 1:3. It rendered English text well but still produced partly garbled Chinese13. On September 8, OpenAI released Images 2.5. The company said it cut generation latency by up to half compared with Images 2.01921. Then, on October 5, OpenAI announced plans to show ads inside the image-generation experience itself27.

The marketing claims have been tested

Coverage of the December launch agreed on the main improvement, which was editing that preserves the rest of the image. Testers did not agree on the speed claim. DataCamp found that full generations were faster but did not reach the fourfold figure. In its testing, the "up to 4x" gain came mostly from targeted edits that regenerate only the changed region14. It also attributed the 20% cost cut mainly to that partial regeneration, noting that per-token prices had not changed from GPT Image 114. Per-token pricing stayed flat, so the savings depend on doing edits rather than full generations, and that is the workload advertising and ecommerce teams are likely to run.

Early users disagreed more sharply. On Reddit, some said Google's Nano Banana Pro was faster and more realistic. Others thought OpenAI's model was clearly better at editing15. One photographer said the model's resolution limit made it unusable for professional work beyond creative direction15. Another user estimated it cost about 3 cents per image against roughly 13 cents for Google's model15. DataCamp framed the release as OpenAI's response to Nano Banana Pro, which had shipped less than a month earlier14. OpenAI itself admitted the results "remain imperfect"11, and Wikipedia's summary notes that some art styles got worse in version 1.512.

The resolution complaint now looks dated. A third-party press release about GPT Image 2.5 says output now reaches up to 4K and claims the platform generates more than 3 billion images a week25. Neither figure appears in the main news coverage, so both should be treated as unverified.

Images 2.5 is built for sharing and campaign work

The September release shows where OpenAI thinks image use is growing. Images 2.5 added Sketch, which turns a drawing made in ChatGPT into a reference for the final image. It also added templates for formats like posters and merchandise, comments placed directly on an image to target edits, and prompt sharing so others can remix an idea with their own photos2419. Gadgets360 reported that templates ask follow-up questions and cover formats such as flyers and product photos18. Adele Li, OpenAI's product lead for images, told Axios the tools are meant to make people "more engaged in the process of creation" instead of passive24.

Prompt sharing and remixable templates are how user-generated content spreads. December's Images tab already included a one-time likeness upload so people could reuse their own appearance in later images11. Together these features turn ChatGPT's image tool into something like a social creation app built into a chatbot.

The developer side targets commercial production more directly. OpenAI split the API into two models at the same price, $8 per million image input tokens and $30 per million output tokens20. Flare is the fast default, suggested for creator content, rapid prototyping and high-volume generation19. Sunburst trades speed for editing precision and is aimed at production-ready campaign creative and polished product images20. In effect, OpenAI is offering one model for bulk social content and another for final ad and product shots. Tech Insider noted that features like inpainting and remixable prompts exist elsewhere. What is new is putting them in a chat app that hundreds of millions of people already open every week23.

OpenAI says Images 2.5 keeps using C2PA metadata and invisible watermarking to mark images made with its tools1917.

The image tool becomes an ad slot

The biggest change came in October. OpenAI announced a visual ad format that will first appear during image generation. It is meant to show product inspiration, people using a product, or the experiences a product makes possible7. Testing starts later this month in the US with an unnamed first group of advertisers16. The ads will reach Free and Go users only. Plus, Pro, Business, Enterprise and Edu accounts stay ad-free2. OpenAI says the ads will be clearly labeled, kept separate from the generated image, and will not influence ChatGPT's answers78.

Reports agree on all of that but describe the placement differently. Some say the ad appears while the user waits for an image to render1. Others put it beside or beneath the finished result58. ExplainX says it is still unclear whether ads appear before, during or after rendering, how many appear per generation and how they will be priced3. That uncertainty matters, because an ad shown during a wait competes for a moment of idle attention, while an ad next to the result competes with the image the user asked for.

Analyst Eric Seufert points out that someone waiting for a picture usually hasn't asked anything commercial. That makes the unit closer to a display ad, targeted using conversion data from advertisers rather than the user's prompt1. OpenAI is building the tools for that. It added data integrations with Hightouch, Tealium and LiveRamp, plus ten attribution partners including AppsFlyer, Triple Whale and Northbeam8. Haus, Measured and WorkMagic are running incrementality tests, and DoubleVerify and Integral Ad Science are piloting brand-suitability checks without access to real conversations3. ContentGrip argues that these measurement connections may matter more for adoption than the new ad format itself9. That reading looks right. OpenAI is putting together the standard machinery of a digital ad business, not just a new ad unit8.

The early results need caution. WeightWatchers reported 15.3% lower cost per acquisition than blended paid search, and Portland Leather said 93% of its ChatGPT ad visitors were new3. These figures come from advertisers, include no spend or sample sizes4, and none of them come from the visual format, which has not launched yet5. For advertisers, the creative demands also change. ProgressiveRobot argues the format favors lifestyle images of products in use over plain product shots5. One detail stands out for users: under earlier help documentation, free users who chose an ad-free experience lost access to image creation, according to PPC Land as cited by Martin Cid4.

The business pressure is clear. Thurrott reports ChatGPT has more than 1.2 billion weekly users and says OpenAI still needs a sustainable business model, which is why ads have spread to free tiers this year7.

OpenAI has left video

Video is the obvious gap. OpenAI announced in March that it would shut down Sora, including both the consumer app and the professional service31. Sora's social app had offered a TikTok-style feed of AI clips, remix tools and "cameos" of users' own likenesses28. It ran into copyright and deepfake concerns soon after launch3234. Disney's three-year licensing deal and planned $1 billion investment are no longer going ahead, according to people familiar with the matter32. A Disney spokesperson described OpenAI's move as a decision to "exit the video generation business"34. Reports linked the shutdown to compute costs and focus ahead of an expected IPO3132.

One view is that OpenAI's move into visual media and user-generated content failed with Sora. A better reading is that OpenAI has moved that strategy into still images, which cost less to run and fit inside ChatGPT. The features that defined Sora (likeness reuse, remixing, trend-driven presets) now appear in ChatGPT Images, attached to an ad product instead of a separate app that burned through compute. Brand-safety concerns have not gone away either. NewsGuard found that Sora 2 produced false-claim videos in 16 of 20 attempts, including a fake Coca-Cola announcement, because its protections covered public figures but not brands35.

For marketers, the result is uneven. OpenAI now competes in image generation and ad creative, and it is starting to sell ads inside its own creative tool. For video, advertisers will have to use other providers such as Runway, Pika and Luma33.