ChatGPT Images 2.5 Adds Sketch Mode and Dual-Model API

OpenAI updated its visual stack with ChatGPT Images 2.5, focusing on faster response times and direct canvas controls. The update addresses a common friction point in AI workflows: adjusting an image without rebuilding the text prompt from scratch.

What changes in execution speed and image quality?

OpenAI claims generation latency dropped by up to 50% compared to Images 2.0. In practical terms, a render cycle that previously took roughly 30 seconds now completes in 15 seconds, accelerating iterative refinement during prototyping tasks.

Coverage from The Verge highlights visual enhancements including more natural lighting and richer textures. The model also demonstrates better retention of core subject characteristics across multiple editing passes.

How does the new Sketch tool alter visual feedback?

Users can activate a canvas overlay inside the chat window by typing @Sketch. Instead of relying solely on text descriptions to describe spatial positions, creators can draw rough shapes with a mouse or stylus directly over the interface.

The feature integrates with regional commenting. Rather than requesting a full regeneration, users can select a specific segment of a generated image—such as changing eye color or adjusting background placement—and prompt changes strictly within that target area.

Which API model fits your workflow?

While mainstream coverage focused on consumer controls, technical documentation analyzed by developer Simon Willison reveals two distinct API endpoints:

• gpt-image-2.5-sunburst: Designed for complex multi-turn editing where retaining subjects from reference photos is critical.
• gpt-image-2.5-flare: Optimized for high-throughput, routine generation where output speed takes priority over complex instruction tracking.

According to data cited in the release, OpenAI's image models have processed over 3 billion image generations across ChatGPT and API endpoints. The separation into specialized API tiers gives engineering teams explicit control over latency costs versus output fidelity in custom pipelines.