OpenAI released ChatGPT Images 2.5 and published its safety documentation. The release covers two models, named GPT-Image-2.5-Sunburst and GPT-Image-2.5-Flare. Editing keeps more of the original detail, and infographic accuracy and layout improve. New options include sketch-based generation, templates and shared prompts. Generation is also faster. OpenAI says the higher realism could make convincing fake imagery of real people easier without safeguards. Requests are checked before generation, and both input and output images are screened by a safety model. Adversarial testing showed fewer violating images reaching users than in the previous version. Neither model crossed OpenAI's high-risk thresholds for biology or cybersecurity. OpenAI still applies its strictest biological-risk blocking as a precaution.

• 1.09% unsafe images shown (Sunburst)
• 1.41% unsafe shown (Flare)
• 1.64% unsafe shown in version 2.0
• Sexual category: 0.52% vs 2.07%