OpenAI tests the new generation image model gpt-image-2, focusing on complex charts and realism

On April 21, it was reported that OpenAI’s upcoming next-generation image generation model has been tested among some ChatGPT users and has appeared on third-party leaderboards under the code name gpt-image-2. The model is expected to be officially released in the coming weeks.
Sources mentioned two key improvements. First, the ability to handle complex charts has been significantly enhanced. For example, precisely placing sticky notes or text at specific positions on a whiteboard, something the old model struggled with. Second, the output no longer has an “AI feel”; previously, overly perfect lighting and skin smoothing in faces were recognized flaws.
The new model is driven by OpenAI’s pressure to grow its user base. ChatGPT’s weekly active users recently plateaued at around 920 million, missing the goal of reaching 1 billion by the end of 2025. Sam Altman issued an internal “code red” at the end of last year, triggered by the popularity of Google’s image model Nano Banana, with improving image generation listed as one of the top priorities in the code red.
OpenAI hopes to replicate the “Ghibli moment” from early 2025, when hundreds of millions of users flooded ChatGPT to transform their and their friends’ photos into Ghibli-style images. With improved text rendering capabilities, images can now be more reliably used in advertising and educational charts. This is one of the directions where OpenAI is trying to outpace Anthropic, which recently surpassed OpenAI in revenue thanks to its programming models.