OpenAI releases DALL·E 3

OpenAI announced DALL·E 3, the third generation of its text-to-image model, integrated directly into ChatGPT rather than offered as a standalone tool. Where earlier image generators required users to craft an effective prompt themselves — a skill that had spawned its own online cottage industry — ChatGPT would act as an intermediary, expanding a short, casual request into a detailed prompt before rendering the image, and letting users refine the result through further conversation rather than by re-writing prompts from scratch.

OpenAI said the model considered each word of a prompt more carefully than its predecessor, producing markedly better adherence to detailed, multi-part instructions, more coherent scene composition, and more legible rendered text within images — a persistent weakness of prior generators. On the safety side, DALL·E 3 was built to decline requests to imitate the style of named living artists and to refuse generating images of public figures identified by name in a prompt, following criticism of earlier image models over both issues.

Access rolled out gradually rather than immediately: OpenAI said DALL·E 3 would reach ChatGPT Plus and Enterprise subscribers in October, with research-lab and API access to follow that autumn. The delayed, staged rollout was itself notable as a departure from the wide, immediate releases OpenAI had used for earlier consumer products.

By folding image generation into the same chat interface used for text, OpenAI collapsed a category distinction — dedicated image tools such as Midjourney and Stability AI’s Stable Diffusion versus general chatbots — that had defined generative media through 2022 and early 2023. It also gave OpenAI a second front, alongside GPT-4’s text and code capabilities, on which to compete directly with Google and the dedicated image-generation start-ups.