OpenAI teases GPT‑Image 2
- OpenAI posted a highly realistic AI‑generated screenshot and hinted at a new image model called “gpt-image-2.” - Early testers and creators say circulating outputs look nearly indistinguishable from real photographs. - The teaser appeared on OpenAI social posts and was reported by The Decoder tonight, driving creator speculation. ( )
OpenAI signaled a new image model on April 21 by posting what looked like a ChatGPT screenshot and saying, “This is not a screenshot.” (x.com) (the-decoder.com) The Decoder reported the model is being referred to as “gpt-image-2” and said OpenAI planned to unveil it in a livestream starting at 12 p.m. Pacific on April 21. (the-decoder.com) (9to5mac.com) Image models turn text prompts into pictures, and OpenAI’s current family already sits inside ChatGPT and the application programming interface, or API, so users can generate or edit images in a chat. OpenAI’s documentation says the latest public model today is GPT-Image-1.5, with GPT-Image-1 and GPT-Image-1-mini also available. (developers.openai.com) (openai.com) OpenAI introduced GPT-Image-1.5 on December 16, 2025, saying it made image generation up to four times faster and rolled out to ChatGPT and the API the same day. The API docs still describe GPT-Image-1.5 as OpenAI’s “latest and most advanced” public image model as of April 2026. (openai.com) (developers.openai.com) The new teaser points at a different use case than older “make me a painting” demos: fake user interfaces, charts, and screenshots that read like real software. The Decoder said early test images show stronger handling of dense text, diagrams, and detailed screen layouts. (the-decoder.com) That matters because OpenAI’s own API now supports computer-use tools that rely on screenshot-based interface interaction, while its image stack already supports multi-turn editing inside conversations. Better screenshot realism and cleaner text would fit directly into those product lines, although OpenAI had not published formal GPT-Image-2 documentation at the time of the teaser. (developers.openai.com 1) (developers.openai.com 2) OpenAI has also been moving users away from DALL·E. Its image-generation guide says DALL·E 2 and DALL·E 3 are deprecated and support is scheduled to end on May 12, 2026, while GPT Image models share the main API surface. (developers.openai.com) Independent reports around the teaser described test outputs as hard to distinguish from photographs, but those claims were still based on limited previews and social posts, not a public benchmark release from OpenAI. Until the livestream and model card arrive, the screenshot itself is the clearest fact: OpenAI used an image to announce a better image model. (the-decoder.com) (9to5mac.com)
