OpenAI launches GPT-Image-2.6 with Flare and Sunburst models

The Flare and Sunburst models

OpenAI has officially shipped the GPT-Image-2.6 model family across its primary platforms, including ChatGPT, Codex, and its developer API. This update introduces two specific models designed for different use cases: Flare and Sunburst. These models are closed-weights systems that process text and image inputs to generate new visual content.

The Flare model is specifically engineered for fast iteration and interactive workflows. It is designed to provide rapid results, with OpenAI stating it can return an image in as little as 6 seconds when the quality is set to low. This makes it a primary choice for developers who need quick feedback loops or high-volume, low-latency outputs.

In contrast, the Sunburst model is built for detailed compositions and high-fidelity editing. It targets users who require stronger consistency across multiple edits and higher rendering detail. While Flare prioritizes speed, Sunburst ensures that complex visual relationships and intricate details remain stable throughout the creative process.

New complexity and quality controls

A significant technical addition to GPT-Image-2.6 is the new complexity parameter. This field determines how much the model plans before it begins the actual rendering of an image. Users can select from three distinct levels: low, medium, and high, with the system defaulting to high if no selection is made.

The complexity levels serve different practical purposes for different types of prompts. Low complexity is ideal for focused edits, such as swapping out a background or changing a specific sign in a photo. High complexity is better suited for dense text, complex layouts, and images featuring complicated relationships between multiple objects.

OpenAI notes that while higher complexity levels can improve results for difficult prompts, they may also increase latency and occasionally introduce unwanted changes. Importantly, adjusting the complexity parameter does not incur a separate fee. It remains distinct from the quality setting, which specifically controls the rendering detail and the overall fidelity of the final image.

GIFs, SVGs, and markup edits

For general users on ChatGPT, the 2.6 update introduces the ability to turn prompts or photos into animated GIFs. These animations can be customized by adjusting the speed and editing individual frames before they are shared. This removes the need for a separate animation step for teams working on social media or product content.

Developers using the Image API gain access to SVG output, which is a vector format that allows icons, diagrams, and UI assets to scale without becoming blurry. Additionally, the update includes markup-guided edits. This feature allows users to mark up the exact region of an image they wish to change, providing much more precise control over the editing process.

The release also includes a feature for users to save and tag reference images of people, pets, and characters. By tagging these references, users can reuse specific subjects in future prompts without having to re-upload the original files every time. This helps maintain character consistency across different generated scenes.

Performance and pricing details

The performance of the new models is a major highlight, as OpenAI reports that Images 2.6 generates images 50% faster than the previous version. This speed improvement comes as the platform sees a massive increase in total volume. According to OpenAI, people are now generating more than 4B images per week across ChatGPT and the API, up from 3B per week just eight weeks ago.

Pricing for the new models remains consistent with the previous version, using a token-based system. For the API, text input costs $5.00 per 1M tokens, while image input costs $8.00 per 1M tokens. The cost for image output is set at $30.00 per 1M tokens for both the Flare and Sunburst models.

While the speed and features are well-defined, OpenAI has not yet published a per-image cost table or specific benchmark scores for the 2.6 release. This means developers will need to calculate their own costs based on the token usage of their specific prompts and the resulting image outputs. The models remain hosted exclusively on OpenAI's infrastructure.