OpenAI Images 2.0

OpenAI on April 21 rolled out ChatGPT Images 2.0, a new image model that can plan before it draws when users switch on “Thinking.” (openai.com) OpenAI said the model improves text rendering, multilingual output, and visual reasoning, and made ChatGPT Images 2.0 available across all ChatGPT plans. “Images with thinking” is narrower: it is limited to paid plans when users select Thinking or Pro models. (openai.com, help.openai.com) In OpenAI’s system card, “Thinking” means the image tool can use the company’s reasoning stack and live web search, generate multiple candidate images, and refine a prompt before producing a final image. OpenAI said that extra capability is meant to turn a basic request into a more researched output. (deploymentsafety.openai.com) Image models turn text prompts into pictures, and one of their persistent weak spots has been putting long, accurate text inside an image. OpenAI is pitching this release around that problem, saying the new model is better at dense text, instructions, and layouts that look more like slides, diagrams, or interfaces than posters. (openai.com, deploymentsafety.openai.com) The timing lines up with a broader OpenAI push toward agents, or software that can keep working across tools and over longer sessions instead of answering one prompt at a time. On April 22, one day after the image launch, OpenAI introduced workspace agents in ChatGPT for Business and Enterprise customers. (openai.com, openai.com) OpenAI said those workspace agents are Codex-powered, run in the cloud, can be shared across a company, and can keep working when the user is away. Enterprise release notes said they are rolling out over the next few weeks to Business and Enterprise workspaces and can connect to apps including Google Drive, Google Calendar, Slack, and SharePoint. (openai.com, help.openai.com) That release followed two more agent-focused updates the same week. On April 15, OpenAI said its Agents Software Development Kit was getting native sandbox execution and a model-native harness for long-running work across files and tools; on April 16, it expanded Codex with computer use, in-app browsing, memory, plugins, and ongoing tasks. (openai.com, openai.com) OpenAI’s safety write-up said the stronger image model also raises the risk of more convincing fake images of real people, places, and events. The company said it added prompt filters, image-input checks, and output screening on top of the safety system it used for earlier ChatGPT image releases. (deploymentsafety.openai.com) The release leaves OpenAI selling images less as a standalone art tool and more as one part of a longer-running assistant. The new model draws the picture, but the product pitch is increasingly about software that can research, plan, edit, and keep going after the first prompt. (openai.com, openai.com, openai.com)