OpenAI Releases GPT Image 2.5 With More Precise, Consistent Editing

Summary

OpenAI has released GPT Image 2.5, positioning precise editing, multi-turn consistency and faster generation as the main upgrades over Image 2. The model is designed to change a specified element while preserving the subject, composition, background and brand appearance, and to carry earlier edits into later rounds without steadily degrading image quality. OpenAI offers two API variants: Flare, the faster default, and Sunburst, which prioritizes control and quality for uses such as advertising and product imagery. According to the Arena results cited in the article, the two variants took the first and second places across text-to-image, single-image editing and multi-image editing, while Image 2 occupied the remaining position in the top three. Relative to Image 2, Sunburst reportedly gained 40, 59 and 81 points in those three categories, while Flare gained 18, 30 and 47 points; Flare also cuts generation latency by up to 50%. The article says both variants were listed at the same prices as Image 2: $8 per million image-input tokens, $30 per million image-output tokens and $5 per million text-input tokens. GPT Image 2.5 also adds Sketch input on mobile and a way to share a generated image together with its prompt. The article argues that these interaction ideas largely follow capabilities already seen in Google’s Nano Banana and other products, so the release is more of a catch-up in workflow than a fundamental invention. Users have demonstrated stop-motion sequences, animation frames generated from sketches and transparent-background assets. Reported weaknesses remain, including malformed hands, high-frequency noise, grain, pixelation, over-sharpening and occasional failures to follow complex instructions. For strict product reproduction, the article says Nano Banana 2 and Pro can still be more reliable.