OpenAI officially released ChatGPT Images 2.5 on September 8, 2026, targeting creative professionals, game artists, and commercial designers. According to the company, users currently generate more than three billion images per week across the platform. The new release focuses directly on common bottlenecks in digital creative pipelines, notably high generation latency and inconsistent visual continuity across iterative steps.
Generation speed is a key technical highlight of the release. OpenAI claims latency has been cut by up to 50 percent compared to Images 2.0, significantly speeding up interactive design sessions. At the same time, the model improves multi-turn consistency. Core reference subjects and distinct character features now remain coherent across repeated conversational prompts rather than shifting unexpectedly during iterative refinement.
The update also introduces refined user interface controls directly inside the chat environment. Users can now draw interactive sketches to guide composition, layout, and perspective before generating an asset. Furthermore, a point-and-click canvas commenting system allows creators to mark specific visual areas and provide localized revision notes, enabling precise touch-ups without modifying unintended parts of an image.
For studio workflows and software developers, OpenAI is providing two specialized application programming interfaces. The gpt-image-2.5-flare model is tuned for high throughput and minimal latency, making it suited for creator content, rapid prototyping, and high-volume image creation. Teams can integrate the endpoint into real-time applications where prompt turnaround time is the primary performance requirement.
Complementing the speed-focused tier, OpenAI introduced gpt-image-2.5-sunburst for deep precision. This second API model focuses on multi-turn fidelity and complex visual adjustments across extended modification sequences. It allows design studios and software builders to construct automated production pipelines where subtle visual details must survive across multiple consecutive image editing operations.