The Story Behind GPT Image 2
For most of the DALL-E 3 era, AI image generation had a famous weak spot: text. Ask DALL-E 3 for "a coffee shop sign that says ‘Open Late’" and you'd often get "Opem Latee" or "Opwn Lqte" — a passable photograph ruined by garbled letters. Multilingual support was even worse: Japanese, Korean, and Arabic prompts produced pseudo-calligraphy that looked vaguely like the script but wasn't real text.
GPT Image 2 — released by OpenAI in 2025 — was the first mainstream model to substantially solve those problems. Internal and third-party benchmarks place its multi-word text rendering accuracy at roughly 99%, and it natively renders 7+ scripts (Japanese, Korean, Simplified Chinese, Traditional Chinese, Arabic, Spanish, French and more) at parity quality. Photorealism took a similar step forward: outputs are routinely indistinguishable from real photography in blind tests.
You may also see GPT Image 2 referred to as ChatGPT Images or ChatGPT Images 2.0. That's not a different product — it's the informal name users gave the model when it shipped inside the ChatGPT interface, where Plus subscribers generate images by simply chatting. The underlying model is the same one available through the OpenAI API and through third-party platforms like GPTImager.
For a deeper read on the ChatGPT integration specifically, see our ChatGPT Images explainer.