An image model designed for continued editing
OpenAI released ChatGPT Images 2.5 on September 8. The company says the model better preserves people and objects from references, renders more natural light and texture, changes only the requested region more reliably and maintains prior edits across longer conversations. It also reports up to 50% lower image-generation latency than Images 2.0.
ChatGPT adds Sketch for drawing a visual reference, templates for common formats such as flyers and product photos, comments placed directly on an image, and prompt sharing. The rollout covers all tiers of ChatGPT, ChatGPT Work and Codex across desktop, mobile and web. Developers receive two API variants: GPT‑Image‑2.5 Flare for quality and speed, and Sunburst for additional precision with longer generation time.
The limits behind “up to 50%”
The speed and quality improvements are OpenAI's own measurements. Results can vary with prompt, resolution, quality setting and edit history. Better reference fidelity does not guarantee perfect identity, product geometry or rendered text. Production teams should run blind comparisons on the same references and separately list what must change and what must remain fixed.
OpenAI says it continues to attach C2PA metadata and invisible watermarks. Metadata can be stripped, so provenance should be evaluated with multiple signals rather than treated as a single definitive detector.