OpenAI introduced ChatGPT Images 2.5 on September 8, 2026, with a stronger emphasis on controlled editing rather than one-shot image generation. The release adds sketch-guided creation, reusable format templates, comments placed directly on images and prompt sharing. OpenAI says the model also improves reference-image fidelity, multi-turn consistency, lighting and texture, while reducing generation latency by up to 50% compared with Images 2.0. These are vendor claims, not independent test results.
The announcement matters because image generation is moving closer to an iterative design workflow: start with a rough idea, refine one element, review it and repeat. That can be more useful for a small business than repeatedly regenerating an entire picture and losing details that were already correct.
What has changed?
The new Sketch feature lets a user draw a rough layout inside ChatGPT and use it as a visual reference. Templates provide starting structures for formats such as posters, merchandise and product photos. Users can also place comments on specific areas of an image to make an edit request more precise.
OpenAI says Images 2.5 is better at changing only the requested element while preserving the subject, composition and visual treatment. It also claims that earlier choices are more likely to survive a series of edits. The full feature description and examples are in OpenAI’s September 8 announcement.
A practical small-business workflow
A local business could use the new tools to prototype a flyer or product visual without treating the first AI output as finished work. A sensible process would be:
- Write the brief first. Define the audience, purpose, dimensions, brand colours, required text and elements that must not change.
- Sketch the composition. Mark where the product, headline and contact area should appear. The sketch communicates layout without requiring polished drawing skills.
- Generate a draft. Use generic or properly licensed source material. Do not upload customer photographs, confidential documents or third-party artwork without permission.
- Edit one variable at a time. Request a background, colour or object change while explicitly listing everything that must remain unchanged.
- Verify outside the model. Check spelling, phone numbers, prices, product details, proportions and required legal wording. Compare any depicted product with an approved reference.
- Export and archive. Keep the final prompt, source permissions, approval record and an unedited reference file.
For example, a hypothetical repair company might sketch a flyer with a machine image on the left and an empty information area on the right. It could then test different backgrounds while keeping the equipment shape and layout stable. This is a suggested workflow, not a report of an actual test.
Availability and developer options
OpenAI says Images 2.5 is rolling out across all ChatGPT tiers on desktop, mobile and web, as well as ChatGPT Work and Codex. Interface availability can still vary during a rollout, so readers should confirm the feature appears in their own account before committing to a deadline.
Developers get two API models: GPT-Image-2.5 Flare, positioned for faster general workflows, and GPT-Image-2.5 Sunburst, positioned for detailed work that can take longer. API costs are usage-based and can change, so teams should check the linked official pricing page rather than rely on an old quoted price.
Important limitations and safety checks
Better consistency does not make an image automatically accurate, original or legally usable. Generated text may still need correction. Real people should consent to edits, and businesses should avoid misleading product depictions, fake endorsements and copyrighted styles or assets they are not permitted to use.
OpenAI’s Images 2.5 system card warns that increased realism can enable more convincing deepfakes. It describes prompt and image safeguards, while also acknowledging that no provenance solution is complete. OpenAI says outputs use C2PA metadata plus invisible watermarking, but businesses should still label AI-generated or materially edited visuals when context could mislead viewers.
The useful takeaway
The most valuable improvement is not simply “better pictures.” It is a more controllable draft-review-edit loop. Teams should test the model on a low-risk asset, record the time and corrections required, and compare the result with their existing design process. Human approval remains essential for accuracy, rights, brand consistency and publication.
