The images arrived without much ceremony. On September 9, OpenAI released ChatGPT Images 2.5, its newest image generation model, and made it available to every ChatGPT, ChatGPT Work, and Codex user on desktop, mobile, and the web. Two API models, GPT-Image-2.5 Sunburst and GPT-Image-2.5 Flare, shipped the same day for developers.
The quiet rollout is notable for what it does not do: it does not gate the feature behind a premium tier or a waitlist. OpenAI has been steadily folding its newest image capabilities into the default experience, treating images as a core feature rather than a separate product. Giving Codex users access at launch signals where the company sees demand: developers who want to generate assets inside a coding workflow.
The new model continues a line that has defined OpenAI’s consumer reach. DALL-E 2 arrived in 2022 and made text-to-image generation a mainstream phrase; DALL-E 3 followed in 2023 inside ChatGPT. GPT-Image-1, released in April 2025, was the first built directly on a GPT model rather than a separate image network. The 2.5 label suggests an incremental step in that direction, with Sunburst and Flare marking variants aimed at different styles or speeds.
The most memorable chapter in that history was not a technical paper but a meme. In the spring of 2025, users flooded social media with images in the style of Studio Ghibli animation, generated at a volume that briefly strained OpenAI’s servers. The episode proved image models could drive mass adoption the way chatbots had, and it put pressure on rivals to match the output.
That competition has tightened since. Google has pushed image generation inside Gemini, xAI’s Grok added its own image model, and Midjourney continues to serve a dedicated base of professional artists. OpenAI’s edge has been distribution: image generation inside ChatGPT reaches hundreds of millions of users without asking them to learn a new tool.
The images release also fits a larger multimodal push. OpenAI has positioned its models as able to read and produce text, images, audio, and video within a single system, and image generation is the most mature of those output modes. Getting it into every account, including Codex, is a step toward making the model a general-purpose tool rather than a collection of separate features.
Text rendering is where image models have most visibly improved. Early systems mangled letters and numbers inside pictures, which limited their use for anything with a label, a sign, or a caption. Each generation has gotten better at typesetting inside images, and that single capability has expanded the practical use cases far beyond art, into marketing materials, diagrams, and interface mockups.
Image generation is also a compute story. Rendering a single high-resolution image can cost an order of magnitude more than answering a text prompt, and the platform bears that cost on every request. OpenAI has managed the economics by moving image work onto its own models and bundling it into subscriptions, spreading the expense across millions of users who generate a handful of images rather than thousands. The tiering implied by Sunburst and Flare, one name suggesting richness and the other speed, is a way to match quality to what a given request is willing to pay.
The legal backdrop remains unsettled. Getty Images sued OpenAI in early 2025 over training data, and a group of artists filed a separate class action. OpenAI has argued that training on copyrighted material is fair use, and the company has not publicly detailed the data behind Images 2.5. Those cases are still moving through courts, and their outcomes could reshape the economics of every image model.
OpenAI has also moved to attach provenance to its output, embedding metadata such as the C2PA standard that records how an image was made. The company has described this as part of its approach to synthetic media, giving platforms and users a way to identify generated content. Watermarking and labeling have become default expectations from regulators and platforms alike, and the images model is one place those expectations are enforced by design rather than policy.
For businesses, the practical question is control and consistency. Brands that want a specific look across thousands of images have pressed for style locking and reproducible outputs, features enterprise buyers mention more often than raw photorealism. OpenAI has tied the Work and Codex access to the same accounts those customers already bill for text and code, folding images into an existing commercial relationship.
The release is unlikely to move the market on its own. Image generation is no longer a novelty that shifts share prices; it is a table-stakes feature every major platform must offer. What the 2.5 rollout does is remove another reason for a customer to leave ChatGPT, which, in a market defined by switching costs, is often the whole game.
For OpenAI, images are also a retention tool. The company has spent the year cutting prices on models like Luna to pull users in, and every new capability folded into the default experience makes a subscription harder to cancel. ChatGPT Images 2.5 is, in that sense, a retention feature wearing a model name.
