A new model of reasoning about composition, it searches the web for context, generates up to eight matching images from a single notification, and translates text into non-Latin scripts with flawless accuracy. It also took first place on the Image Arena leaderboard within 12 hours of launch, with the highest rating ever recorded.
Two years ago, asking ChatGPT to produce a visual was like sending a poster to an intern with a glue stick and a head injury. You can ask for a clean design and find “leftover intelligence” scattered across the picture, along with three new words that seem to have been invented during a software malfunction.
The images look AI-generated in a way that has become a rare cultural epitome: almost right, obviously wrong, and instantly recognizable as fake.
The leap is important. Text rendering has been a persistent, embarrassing weakness for AI graphics developers since the DALL-E first turned heads in January 2021, the prototype model. incorporated at that time as an interesting curiosity.
Graphics 2.0 claims nearly 99% accuracy in rendering text across all languages and scripts, including Japanese, Korean, Chinese, Hindi, and Bengali. If that figure works in independent testing, it bridges the gap between “an impressive AI demo” and “a tool a graphic designer can actually use in production work.”
The structural changes that make a model different, if not better, are what OpenAI calls “thinking skills.” Images 2.0 is the company’s first image model to integrate its O series imaging architecture.
Before generating a pixel, the model researches the data, organizes the composition, reasons about the spatial relationships between the elements, and can search the web for real-time context.
It is, in OpenAI implementation, not a rendering tool but a “a companion of the visual imagination.”
This is my cat turned into a comic strip with ChatGPT.
Actually, this is seen in two ways of access. Express mode is being sent to all ChatGPT users, including free accounts, and brings key quality improvements: better text, sharper editing, richer layouts.
Think mode, which enables web searches, multi-image stacking, and output verification, is limited to Plus ($20/month), Pro ($200/month), Business, and Enterprise subscribers.
The distinction is commercially important. The power of thought, where most of the quality premium resides, resides behind a paywall. Free users get better photos; paying users get the images that the model thought of.
The power of multiple images is the feature most likely to revolutionize professional workflows. A single input can now generate up to eight images that preserve the character and continuity of the object throughout the set.
That means a designer can produce a family of social media assets, a series of children’s books, or a series of storyboard frames from a single command, with a consistent visual identity throughout.
Previously, each image had to be individually stitched and assembled by hand. For marketing teams and content creators, that’s a logical reduction in production friction.
Integration into The CodexOpenAI’s coding environment, is a fairly loaded motion. Developers and designers can now generate UI images, prototypes, and virtual assets within the same agent workspace they use for code, slides, and browser automation, using a single ChatGPT subscription.
An image model is no longer an independent product; a capability embedded in the extensive OpenAI platform, which is not only competitive with Midjourney and Google Nano Banana 2 in quality but with Canva and Figma in workflow integration.
The benchmark performance is impressive. Within 12 hours of launch, Images 2.0 took first place in Photo Arena leaderboard in all categories, with 1,512 points, +242 points more than the second place model, Google’s Nano Banana 2.
For most of the year 2026, OpenAI and Google have been trading at the top position within a tight range; Images 2.0 broke quickly.
DALL-E 2 and DALL-E 3 are deprecated and deprecated on 12 May 2026. GPT-Image-1.5, released in December 2025 as an intermediate upgrade, remains accessible via the API through legacy integration but is no longer the default model.
OpenAI has not disclosed the architecture of Graphics 2.0, describing it only as a “standard model” or “GPT for graphics” and refusing to specify whether it uses a distributed, automatic, or hybrid method. The API model identifier is gpt-image-2; the API is expected to be open to developers in early May 2026.
Token-based pricing is $8 per million tokens for image input, $2 for stored input, and $30 for image output, with per-image costs typically ranging from $0.04 to $0.35 depending on the speed and complexity of the image. Output resolution is up to 2K.
The information cut-off is December 2025, which presents an effective boundary: the model cannot accurately provide events, people, or products that appeared after that date without supplementing its internal information through a live web search.
The security structure of this model includes content filtering, C2PA metadata for its appearance, and what OpenAI described in the press conference as continuous monitoring, a point that the company emphasized a lot, given the growing control of artificial media tests and the use of AI image generators in deepfakes, scams, and illegal images.
The most important question raised by images 2.0 is not about quality. The technological gap between AI-generated and human-generated images has been narrowing for years; this model reduces it further.
The question is about what happens when a tool is no longer a novelty but an infrastructure, when image production is the automatic ability of every coding environment, every conversation area, and every business production organization, and when the difference between “human design” and “information production” becomes something only metadata can verify.
OpenAI, on the other hand, seems to be betting that the answer is scale: more, faster, better, cheaper images, everywhere. When we put together first compiled by DALL-E Five years ago, the results of the model were an interesting surprise. They are now a production asset.
The time when AI-generated images were obviously AI-generated is over. The next thing depends on whether the guardrails can keep up with the skill.
