OpenAI has launched ChatGPT Images 2.5, its new state-of-the-art image model — an upgrade to the pipeline through which people already create more than 3 billion images every week.
The Model Improvements
- Sharper details — more natural lighting and richer textures; better at preserving subjects from reference photos so “distinctive features are more likely to carry through”
- Precise editing — edits change only what’s requested while the rest of the image stays intact, even with complex subjects and backgrounds
- Multi-turn consistency — earlier changes stay consistent through longer conversations; each edit builds on previous work without quality degradation
- 50% lower latency vs Images 2.0
- Better instruction intelligence — more accurate real-world content, complex layouts including transparent backgrounds, and more faithful style rendering
New Product Features
- Sketch — draw directly in ChatGPT as a reference; type “@Sketch” and turn rough drawings into complete images
- Templates — starting points for popular formats like posters, flyers, and product photos
- On-image comments — place feedback directly on images for focused editing
- Shareable prompts — share images with the prompt that created them, so others can remix with their own photos (the viral “‘80s portrait” prompt is the showcase example)
The API Models
Two new models ship for developers:
- GPT-Image-2.5 Flare — the default: higher quality than GPT-Image-2 at 50% lower latency. Manus reported it delivers images at “two to four times the speed of GPT-Image-2” with improved transparent-background generation.
- GPT-Image-2.5 Sunburst — premium tier for production-ready campaign creative and polished product imagery, trading longer generation times for tighter edit control.
Early partners include Higgsfield AI (“understands what not to change”), Adobe (integrating the models into Firefly), Runway, and Manus.
Safety
The system builds on existing safeguards with prompt and image checks, plus C2PA metadata and invisible watermarking for provenance — details in the published system card.
Context
The launch extends a busy September for OpenAI — arriving five days after GPT-6 Astra and days after the Navier–Stokes announcement. In the image generation race, it counters Google’s Gemini image advances, Nano Banana Pro on Higgsfield, and the open-source FLUX family — with OpenAI betting that speed plus editing precision (rather than raw artistic quality alone) is what production workflows actually need.