OpenAI Launches ChatGPT Images 2.5 with 50% Lower Latency, In-Chat Sketching, and Canvas Annotation

IT Home, September 9 news: OpenAI officially announced today (September 9) the release of ChatGPT Images 2.5, which reduces image generation latency by up to 50% and introduces useful features such as Sketch references, a template library, and in-image comment annotations, making it the company's most advanced image generation model to date. Compared to the Images 2 model released in April 2026, Images 2.5 focuses on reducing image generation latency, cutting latency by up to 50% while improving detail sharpness, natural lighting, and texture richness. In terms of features, ChatGPT Images 2.5 introduces the Sketch feature. Users can invoke this feature by typing "@Sketch", allowing them to draw sketches directly within the conversation and then supplement them with style and detail descriptions for the model to convert the sketch into a complete image. The product also adds templates for common creative formats such as posters and merchandise, lowering the barrier to creation. Users can directly add annotations on the image to focus on specific areas for modification, and can include the original prompt when sharing images. The in-image commenting feature allows users to mark locations on an image and issue commands such as "delete this area" or "change this to blue"; the model will only modify the specified area while keeping the rest unchanged. Across multi-turn editing sessions, the model better maintains image consistency and quality, preventing detail degradation after multiple edits. In terms of API models, OpenAI offers two models: GPT-Image-2.5 Flare: aimed at most API image generation applications, suitable for scenarios such as creator content, social content, product experiences, visual search, rapid prototyping, and high-volume batch generation. GPT-Image-2.5 Sunburst: geared toward workflows with higher demands for editing control. OpenAI recommends it for scenarios such as production-grade marketing assets and retouched product photos, though its generation time is longer. The API supports multiple resolutions, including 1024×1024, 2048×2048, 3840×2160 (4K landscape), etc., with no single side exceeding 3840 pixels, and a total pixel count ranging between 655,360 and 8,294,400. In terms of pricing, IT Home attaches the pricing for both models as follows: Model Modality Input Cached Input Output gpt-image-2.5-sunburst Image $8.00 $2.00 $30.00 Text $5.00 $1.25 - gpt-image-2.5-flare Image $8.00 $2.00 $30.00 Text $5.00 $1.25 - In terms of safety and compliance, Images 2.5 continues existing safety mechanisms, performing checks on prompts and generated images to prevent harmful outputs. All outputs embed C2PA metadata and invisible watermarks to facilitate the identification of images generated by OpenAI tools and combat unauthorized reuse or deepfakes.