GPT Image 2.5
A next-generation AI image generation and editing model introduced by OpenAI. It demonstrates strong performance on inpainting tasks by deeply coupling language understanding with image editing, enabling more natural content blending and semantically aware editing.
Timeline (last 90 days)
GPT Image 2.5's price rises steeply with quality tier, and the highest-quality renders are significantly more expensive than low- and medium-tier drafts.
In single-character scenes, GPT Image 2.5 has a clear advantage in keeping faces recognizable under different lighting, backgrounds and artistic styles.
GPT Image 2.5 can take up to 16 reference images in image-to-image tasks, with latency reduced by about 50% compared with the previous generation.
GPT Image 2.5 is conspicuously absent from the image generation rankings on that list, and multiple commentators pointed out this major omission.
In the documentary portrait task, GPT Image 2.5 independently added a red fez hat to the elderly subject, which was not requested in the prompt.
GPT Image 2.5 costs $0.0326 per image at the high quality tier, with an output resolution of 1536x864.
The final score across 8 head-to-head tests was OpenAI leading 3 to 2, with 3 draws.
The chat completions endpoint of GPT Image 2.5 rejected these image models during testing, requiring calls via POST /api/v1/images.
When editing with GPT Image 2.5, reference images are passed via the input_references field, supporting URLs or base64, with a maximum of 16 images.
GPT Image 2.5's Flare version is priced the same as Sunburst at every tier, but is nearly twice as fast.
7 more timeline events
All Facts (12)
GPT Image 2.5 can take up to 16 reference images in image-to-image tasks, with latency reduced by about 50% compared with the previous generation.
50%UnverifiedGPT Image 2.5's price rises steeply with quality tier, and the highest-quality renders are significantly more expensive than low- and medium-tier drafts.
50%UnverifiedIn single-character scenes, GPT Image 2.5 has a clear advantage in keeping faces recognizable under different lighting, backgrounds and artistic styles.
50%UnverifiedWhen editing with GPT Image 2.5, reference images are passed via the input_references field, supporting URLs or base64, with a maximum of 16 images.
50%UnverifiedThe chat completions endpoint of GPT Image 2.5 rejected these image models during testing, requiring calls via POST /api/v1/images.
50%UnverifiedGPT Image 2.5's Flare version is priced the same as Sunburst at every tier, but is nearly twice as fast.
50%UnverifiedIf aspect_ratio is not specified in the first call to GPT Image 2.5, it may return 3:2 instead of the expected 16:9.
50%UnverifiedThe codename for GPT-Image-2.5 is Sunburst.
50%UnverifiedThe Along the River During the Qingming Festival scene generated by Luna Max over an hour and a half proactively rendered characters, roof textures, and shop fronts as textures displayed in the 3D scene, making full use of GPT-image's image generation capabilities.
50%UnverifiedFor users who have already subscribed to a commercial service, the marginal cost of using GPT Image 2.5 for inpainting is nearly zero.
50%UnverifiedClosed-source models such as GPT Image 2.5 are currently difficult to use directly in ComfyUI unless you write your own nodes.
50%UnverifiedA creator on Reddit stated that using GPT Image 2.5 for inpainting feels like cheating.
50%