Just two years ago, asking ChatGPT to draw a Mexican restaurant menu meant getting something with dishes like “burrto,” “margatas,” and “enchuita.” AI image generators traditionally handled text awkwardly — and that was their most noticeable weakness. On April 21, 2026, OpenAI introduced ChatGPT Images 2.0, and it seems that this problem is gone.
Why AI generators previously could not write text
To understand what changed, it is worth knowing the cause of the old problem. Most previous image-generation models — including DALL-E — worked on diffusion models. Their principle of operation: reconstruct an image from “noise,” gradually restoring structure.
Text in an image occupies only a small portion of the pixels. The algorithm learned general patterns — and simply did not pay enough attention to letters. The result: “burrto” instead of “burrito,” Cyrillic curls instead of real words, pseudo-hieroglyphs instead of Japanese characters.
Researchers had long been looking for alternatives. Autoregressive models — the ones that “predict” an image gradually, similar in principle to large language models — showed better results with text. OpenAI did not disclose exactly which architecture underlies Images 2.0, but the results speak for themselves.
What ChatGPT Images 2.0 can do — the full list of capabilities
Text in images — the main new feature
Images 2.0 generates readable, correctly written text even in complex compositions: restaurant menus, magazine covers, advertising banners, UI mockups, infographics, educational diagrams. Fonts, hierarchy, alignment — the model reproduces all of this with a level of precision that previously could only be expected from a designer.
[H3] “Thinking” and self-checking
The model received so-called “thinking capabilities” — features that were previously the domain of text models. Images 2.0 can:
- Search for up-to-date information on the internet before generation
- Generate multiple images from a single prompt
- Check its own results and correct mistakes
This explains why generating complex objects takes several minutes rather than seconds. But the result — a marketing banner or a multi-panel comic — may be ready to use immediately.
Consistent image series
Images 2.0 can generate up to eight related images from a single prompt while preserving “character and object continuity” — meaning that characters, objects, and style remain the same from frame to frame. This opens possibilities for:
- Storyboards and comics
- Step-by-step instructions with images
- Series of advertising materials in a single style
- Educational content with sequential illustrations
Multilingual support
One of the most important changes for non-Latin languages. Images 2.0 now correctly reproduces text in Japanese, Korean, Chinese, Hindi, and Bengali — not just as a translation, but as natively embedded text in the design. This is especially important for Asian markets, where the Latin alphabet is not the standard.
Flexible formats and resolution
The model supports aspect ratios from 3:1 (wide banner) to 1:3 (vertical smartphone format), as well as generation at resolutions up to 2K. This makes it suitable for real production, not just demonstrations.
Comparison: Images 2.0 versus previous generators
| Capability | DALL-E 3 (2024) | Images 2.0 (2026) |
|---|---|---|
| Text in the image | ❌ Often unreadable | ✅ Readable, accurate |
| Series generation | ❌ One frame | ✅ Up to 8 related frames |
| Internet search | ❌ None | ✅ Available |
| Non-Latin languages | ⚠️ Partial | ✅ JP, KR, CN, HI, BN |
| Resolution | Up to 1024px | Up to 2K |
| Aspect ratios | Limited | 3:1 to 1:3 |
| Self-checking | ❌ None | ✅ Available |
Who this is actually useful for
ChatGPT Images 2.0 is not only a tool for artists and designers. Thanks to solving the text problem and introducing “thinking,” the model becomes a practical tool for a much wider audience.
Marketers and content managers can generate publication-ready banners, social media covers, and advertising materials without involving a designer to fix text.
Educators get the ability to create educational diagrams, infographics, and illustrated step-by-step instructions with correct labels.
Developers via the API (gpt-image-2) can automate image generation with text for their products — menus, product cards, UI mockups.
Bloggers and media outlets — including those writing about technology, gadgets, and artificial intelligence — can quickly create unique illustrations for articles.
Limitations and what the model still cannot do
OpenAI honestly points out the current weaknesses of Images 2.0:
- Cropping issues in complex compositions
- “Hallucinations” — the model may invent details
- Complex charts and diagrams with precise data still need refinement
- Very dense textures and superscript-level details may come out with artifacts
- Precise editing of existing images is still limited
In addition, the model’s knowledge base is cut off at December 2025. This means: if the generation requires up-to-date data (for example, the logo of a new company or a depiction of a recent event), the result may be inaccurate.
How to get access and how much it costs
Images 2.0 has been available since April 21, 2026 through the “Images” tab in ChatGPT. The access structure:
- Free users — basic access to Images 2.0
- Paid users (ChatGPT Plus, Pro, Business) — expanded features, including “Thinking” mode and higher resolution
- API — the model is available as gpt-image-2, with pricing depending on output quality and resolution
- Codex — support for Images 2.0 is also built into the tool for programmers
What this means for competitors
OpenAI is not the only company solving the text problem in AI images. In February 2026, Google released Gemini 3 Pro Image with similar capabilities for dense text. But according to early testers, Images 2.0 outperforms the competitor in reproducing UI elements, screenshots, and sequences of related images.
Midjourney and Stable Diffusion still remain stronger in artistic generation and stylized images. But Images 2.0 is clearly targeting a different segment — practical content production rather than digital art.
In brief: the key things about ChatGPT Images 2.0
- Main new feature: readable, accurate text in images of any complexity
- Thinking mode: web search, self-checking, series of up to 8 images
- Languages: Japanese, Korean, Chinese, Hindi, Bengali
- Formats: from 3:1 to 1:3, up to 2K resolution
- Access: free through ChatGPT; advanced functions — for paid users; API — gpt-image-2
- Knowledge base: December 2025
Article prepared by the TechVisor team — practical IT media for people.
