OpenAI has introduced ChatGPT Images 2.5, a new image-generation release focused on sharper visual details, more precise editing and faster creation. The company says the update improves the way the model handles reference images, complex instructions and a sequence of refinements, while adding new creation tools inside ChatGPT and two related API models for developers.
For everyday ChatGPT users, the update is intended to make image work feel less like a one-shot prompt and more like an iterative creative process. Images 2.5 is designed to preserve subjects, composition and visual direction more reliably as a user asks for changes. OpenAI says generation latency is up to 50 percent lower than Images 2.0, which should shorten the loop between an idea, a first result and subsequent refinement.
Editing with more control
The release centres on targeted edits. OpenAI says Images 2.5 is better at changing the requested element while retaining the rest of an image, even for complicated subjects and backgrounds. It also aims to maintain consistency through longer conversations, so earlier choices are less likely to disappear when a user adds a later instruction. That is useful for product imagery, campaign concepts, visual prototypes and other work where a small requested change should not require rebuilding an asset from scratch.
OpenAI is also introducing comments on images in the mobile experience. Users can open a generated image, add a comment where they want a change and describe the adjustment. The feature provides a more direct way to communicate a visual edit than repeatedly translating a location into prose. Image prompts can also be shared, allowing another person to begin from the same creative instruction and adapt it with their own details.
Sketch and templates
Sketch is a new input method that lets a person draw directly in ChatGPT and use the result as a guide for generation. A quick outline of a room, outfit or layout can be paired with a written description of the desired style. The value is not artistic polish: a rough drawing can supply composition or placement information that is awkward to express in text alone. OpenAI says the feature is accessed by choosing Sketch from the message composer on supported mobile experiences.
Templates provide another starting point. Rather than beginning with an empty prompt, a user can select a format such as a poster or merchandise design and supply the message, style and visual elements. Templates are meant to reduce setup effort for common image tasks. OpenAI notes that template availability is not yet the same in Work mode, so teams should check their product surface before depending on a particular creation flow.
API models for developers
Developers receive two GPT-Image-2.5 options. Flare is positioned as the default for most applications, combining the quality and editing improvements with lower latency for high-volume use cases such as social content, product experiences, visual search and rapid concepting. Sunburst is aimed at premium workflows that justify longer generation times in exchange for more detailed control. The split gives builders a way to choose between general throughput and a more deliberate creative process.
OpenAI highlights better handling of natural lighting, textures, transparent backgrounds and complex layouts, along with stronger adherence to visual style. These are practical claims rather than a guarantee for every prompt. Teams integrating image generation should test representative briefs, references and brand constraints, especially where an output will be used in customer-facing material. They should also account for the review steps their own publishing process requires.
Availability and safeguards
Images 2.5 is rolling out across ChatGPT, ChatGPT Work and Codex on desktop, mobile and web, and the new API models are available to developers. Existing image-generation limits remain unchanged. OpenAI says it continues to use prompt and image checks, C2PA metadata and invisible watermarking to help identify images made with its tools. Those measures sit alongside the product improvements, not outside them.
The release broadens what users can create and revise without changing the basic discipline required for responsible image use. Faster generation and more accurate edits can make creative work more productive, but organisations still need clear ownership, review and usage rules. For individuals, the most useful first step is likely a small project that tests reference-led editing or Sketch before moving the new model into a larger visual workflow.
This announcement matters because it turns a technical capability into something teams can adopt in regular work. The release combines a clear product change with practical controls, while leaving organisations responsible for deciding where it belongs in their workflows. Users should confirm availability, relevant limits and governance settings in their own account before treating any feature as universally enabled.
As with any new AI capability, the most useful adoption path is gradual. Start with a bounded task, compare the result with an existing process, and document the controls, costs and review points that remain necessary. That approach makes it easier to distinguish a meaningful improvement from a promising demonstration and gives teams evidence for their next decision.