OpenAI has introduced the GPT-Image-2.6 image model family in Flare and Sunburst variants. The former is optimized for faster generation and editing, while the latter delivers more faithful and detailed results. The models are available in ChatGPT, Codex, and through the OpenAI API.
GIFs, vector export, and annotation-based editing
The API now supports creating a looping animated GIF from a text prompt or source image. The animation can contain up to 16 frames. Export to the SVG vector format is also available, remaining sharp when resized.
For edits, users can add drawn annotations that specify both the location and nature of the desired change. The complexity setting in the API controls how much planning the model performs before rendering.
How the models performed in generation and editing
Flare and Sunburst took the top two spots in the Artificial Analysis rankings for both text-to-image generation and image editing. In anonymous comparisons, they outperformed generation 2.5; the quality rankings are based on votes in direct side-by-side output comparisons.
Sunburst led in seven of the ten image generation categories tracked. Its strongest areas included architecture, user interface design, text rendering, and prompts requiring reasoning. Flare ranked first in the consumer use and marketing and advertising categories. Human anatomy remained a weakness for both variants, particularly hands, faces, and dynamic poses.
One of the 2.6 models led in six of the seven editing categories. Sunburst ranked first in areas such as identity preservation, while Flare led in reasoning-based editing and composition. Microsoft MAI-Image-2.7-Preview retained first place in scene and style editing. OpenAI also states that Sunburst's image consistency across edits has improved at medium and high quality settings.
Flare is faster, while Sunburst needs more time
In Artificial Analysis tests, Flare's median generation time fell from approximately 55 to 30 seconds, while editing time dropped from 58 to 32 seconds. The measurements used maximum quality, 1024 × 1024 PNG output, and the API's default high complexity setting.
Artificial Analysis measured Sunburst at approximately 135 seconds to create an image from text and 120 seconds to edit one. In paired tests, it was slower than its predecessor.
Availability in ChatGPT, Codex, and the API
The GPT-Image-2.6 family is available across all access tiers in ChatGPT, Codex, and through the OpenAI API. Both variants cost $30 per million output image tokens in the API, the same as generation 2.5; Artificial Analysis estimates the price of a 1024 × 1024 image at maximum quality at approximately $0.21, with actual costs depending on settings and token usage.
