OpenAI has unveiled ChatGPT Images 2.0, an update to its image generation tool built on the new gpt-image-2 model. The release comes a few months after GPT-Image-1.5 in December 2025, bringing some genuinely useful features alongside the usual marketing polish.

The headline addition is a “thinking” mode: the model reasons before generating, can search the web during the process, and can produce up to eight coherent images from a single prompt. OpenAI sees this as useful for creating comic strips, social media graphics, or variations on a visual concept.

What actually matters

Text rendering in images improves noticeably, including non-Latin scripts like Japanese, Korean, Hindi, and Bengali. Supported aspect ratios range from ultra-wide (3:1) to narrow vertical (1:3), with output up to 2K resolution.

The reasoning features are exclusive to Plus, Pro, and Business subscribers. Everyone else gets real improvements, but more modest ones: better overall visual quality, improved text rendering, denser compositions. The model is also available via API under the name gpt-image-2.

Worth noting: during a press briefing, OpenAI declined to answer questions about the architecture underlying the new model. Training data cuts off at December 2025, which could show up as weaker results on prompts tied to more recent events.

The model had been circulating for weeks on LM Arena under the codename “duct tape” before the official announcement.