The new ChatGPT Images 2.0 can generate people, screenshots, and text (though as you can see, the latter still has issues)
OpenAI has released ChatGPT Images 2.0. This is the biggest update to the image generation system since the very first version of ChatGPT. The new model doesn’t just draw pictures from descriptions — it first “thinks,” can search for information online, and delivers results in up to 2K resolution. The release took place on April 21, 2026, and is available to all ChatGPT users, even free ones.
Image Creation Modes in ChatGPT Images 2.0
According to the OpenAI blog, the new model operates in two modes, each designed for a different purpose. Instant is a fast mode optimized for speed. Before launch, OpenAI tested it on the LMArena platform under the codename “duct tape.” It produces quality images in seconds and is available to all ChatGPT users.
Thinking mode works slower but “reasons” before generating an image. This allows it to maintain character consistency across multiple frames, search for data online, and create up to eight images per request. Thinking is available to Plus, Pro, and Business subscribers.
The difference between the modes is fundamental. Previous models generated images as a one-time action: one request — one result. Thinking turns image creation into a structured process where the model first analyzes the task and only then draws. This opens the door to creating manga, storyboards, and image series with the same characters.
An example of a unique comic created by ChatGPT Images 2.0
Photo Editing in ChatGPT
The most notable change lies in the very approach to working with images. OpenAI no longer treats generation as a one-time “request — response” action. Now it’s a dialogue.
Users can refine images in the process: zoom into fragments, change composition elements, adjust style — all without starting from scratch. The model remembers the context of previous edits and develops the result iteratively.
During the demonstration, the system generated eight summer outfit variations from a single uploaded photo. In another example, it analyzed social media reactions to previous test models, visually summarized the results, and even created a QR code linking to ChatGPT.
This demonstrates the main idea behind the update: ChatGPT Images 2.0 combines reasoning, information search, and design into a single workflow. Previously, this required several different tools.
Eight different summer outfits for one girl
The Best AI for Generating Images with Text
One of the most frustrating problems with all AI image generators has been unreadable text. Inscriptions on posters looked like gibberish, and signs on generated streets resembled an alien alphabet. Images 2.0 attempts to solve this problem.
A doctor’s note made via ChatGPT. Are we entering an era of mass forgery?
OpenAI claims a “quantum leap” in text rendering. The model now handles small fonts, UI elements, dense layouts, and even iconography. According to the company, readable typography works even in complex compositions like magazine covers and infographics.
Hi-news newspaper as imagined by ChatGPT Images 2.0
Support for non-Latin alphabets has also improved — Japanese, Korean, Chinese, Hindi, and Bengali. This was a long-standing weakness of generative models that were primarily trained on English-language data.
A message from Tim Cook in Telegram
However, it’s too early to expect perfect results. Journalists who tested the model on launch day noted that it still makes mistakes with accurate reproduction of logos and brands.
This is not a screenshot but an image created by ChatGPT Images 2.0
Image Formats in ChatGPT
On the technical side, Images 2.0 has received several important improvements. The model supports flexible aspect ratios, from ultra-wide 3:1 to ultra-vertical 1:3. This covers virtually all common formats:
- banners and presentations;
- mobile stories and vertical video formats;
- posters and print layouts;
- square posts for social media.
Maximum resolution has increased to 2K, and you can get up to eight images per request. For developers, the model is available via API under the name gpt-image-2.
A snake in 1:3 format
IMPORTANT: DALL-E 2 and DALL-E 3, OpenAI's previous image generation models, will be discontinued on May 12, 2026. Images 2.0 effectively becomes their replacement, although it is built on an entirely different architecture. Previous versions of DALL-E functioned as a separate tool that ChatGPT accessed "externally." Now image generation is built into the model itself.
Why OpenAI Is Betting on Image Generation in ChatGPT
The text models of leading labs — OpenAI, Google, and Anthropic — are gradually converging in quality. Differentiating on text alone is becoming increasingly difficult. OpenAI, judging by this release, sees image generation as the next competitive frontier.
