OpenAI Launches GPT-Image 2.5 With Two New API Models: Sunburst and Flare

OpenAI has released ChatGPT Images 2.5, an update to its image generation stack that ships as two distinct API models: gpt-image-2.5-sunburst and gpt-image-2.5-flare. The update targets three specific improvements, according to OpenAI: better instruction-following across multiple editing turns, faster response times, and improved preservation of subjects from reference photos.

Two models, two use cases

Rather than a single model update, OpenAI split the release into two variants with different tradeoffs. Per OpenAI's own developer documentation, Sunburst is recommended "for workflows where editing precision matters most," while Flare is positioned for "fast, high-quality everyday image generation." This suggests Sunburst is the higher-fidelity, likely slower and possibly more expensive option, while Flare trades some precision for speed — though OpenAI has not published a side-by-side benchmark comparing the two.

Pricing for both models has not yet been disclosed in the source material reviewed.

Scale claim

OpenAI claims its image models — spanning ChatGPT Images and the GPT-Image family in the API — have collectively generated more than 3 billion images to date. This is a company-reported figure and has not been independently verified.

Early testing

Independent developer Simon Willison tested the release by updating his openai_image.py command-line tool to support passing in reference images alongside a prompt. In one example, Willison fed the tool a line chart image and the prompt "add a raccoon scientist studying the chart thoughtfully" using gpt-image-2.5-sunburst. The output preserved the original chart's data, axis labels, and layout while inserting a cartoon raccoon in a lab coat at a desk in the foreground — a practical demonstration of the model's claimed improvement in reference-image fidelity during edits.

No formal benchmark scores (e.g., prompt-adherence or image-quality metrics) were included in OpenAI's announcement or in third-party testing referenced here.

What this means

This release continues OpenAI's pattern of shipping specialized model variants rather than a single general-purpose update — mirroring how it splits reasoning and non-reasoning text models. For developers, the practical decision is now which variant fits a given workflow: Sunburst for iterative editing pipelines where subject consistency across turns matters, Flare for high-volume, latency-sensitive generation. The 3-billion-image claim signals that image generation has become a meaningful production workload for OpenAI, not just a demo feature, which likely explains the investment in splitting the model line by use case rather than issuing a single incremental update. Until OpenAI publishes pricing and benchmark data, teams evaluating a migration should test both variants directly against their existing editing and generation prompts.