OpenAI accidentally leaked GPT Image 2: new neural network draws YouTube interfaces better than the actual YouTube

In early April, users noticed three unknown models on the Chatbot Arena platform that generate images, and apparently this is the yet-to-be-announced GPT Image 2 — the next generation of OpenAI's generative neural network. Shortly after, the models were removed, but screenshots and examples of their work have already spread across the internet.

Judging by the published images, the new model does what generative networks previously found difficult. First, text rendering on images — in one test, the model correctly displayed a clock face with hands showing the correct time and reproduced text on a sticker, while the Nano Banana Pro model got the hand positions wrong on the same prompt. Second, the neural network handles complex interfaces — users have shared a generated version of YouTube's homepage with video previews, readable headlines, and the correct grid structure. The image doesn't just look like a stylization — it's almost an exact copy of the real interface.

The leak also revealed weaknesses — the neural network still suffers from factuality issues. Additionally, users noticed strangely behaving censorship: on a generated world map, the country of Niger turned into "Ciger" — clearly an aggressive filter kicking in that replaces unwanted letter combinations.

The exact release date is unknown — the previous generation (GPT Image 1.5) the company released in December 2025, and if OpenAI sticks to a six-month update cycle, the announcement of GPT Image 2 could happen in summer or fall 2026.

What do you think — should OpenAI rush the release, or spend more time fine-tuning the model to fix the factuality and censorship issues? Share your thoughts in the comments.