GPT-Image-2.5 Sunburst: OpenAI’s Premium API Image Model Explained
GPT-Image-2.5 Sunburst is the premium-tier image generation model released by OpenAI on September 8, 2026, alongside its faster sibling GPT-Image-2.5 Flare. Where Flare optimizes for speed and volume, Sunburst is built for production workflows that demand exacting control — campaign creative, polished product imagery, and multi-step visual projects where every edit must land precisely. It is part of the broader ChatGPT Images 2.5 family.
What Makes Sunburst Different
The defining characteristic of Sunburst is tight control across iterative edits. In multi-turn generation sessions, each new instruction builds on prior work without degrading quality — elements you didn’t ask to change stay intact, and earlier edits carry through the conversation. This makes it the right tool when a campaign requires ten rounds of refinement rather than a single prompt-to-output pass.
The trade-off is generation time. Sunburst takes longer per image than Flare, which is why OpenAI positions Flare as the default for most applications. Sunburst is the deliberate choice for teams that can wait an extra moment in exchange for production-ready output.
Safety and Adversarial Performance
Sunburst is also the safer of the two API models by the numbers. OpenAI published a system card on September 8, 2026 detailing a layered safety stack: LLM-based policy checks refuse violative requests before generation begins, and a safety reasoning model inspects both inputs and outputs. In automated adversarial evaluations, Sunburst produced unsafe generations at a 1.09% rate — compared to 1.41% for Flare and a 1.64% baseline for Images 2.0. Neither model crosses Bio High or Cyber High thresholds under OpenAI’s Preparedness Framework.
Every image carries C2PA provenance metadata through the C2PA Conformance Program and Google DeepMind SynthID invisible watermarking, covering outputs from ChatGPT, Codex, and the API alike.
Pricing
Sunburst is priced identically to Flare, per the OpenAI API pricing page:
| Token type | Standard | Cached |
|---|---|---|
| Image input | $8.00 / 1M | $2.00 / 1M |
| Image output | $30.00 / 1M | — |
| Text input | $5.00 / 1M | $1.25 / 1M |
You pay the premium-model rate regardless of which variant you call — the differentiation is quality and control, not cost.
When to Choose Sunburst
Sunburst suits workflows where output goes directly into a deliverable: brand campaign assets, polished e-commerce product shots, or any scenario where a creative director will scrutinize every pixel. If you need high-volume throughput or rapid prototyping, Flare is faster and equally priced.
For context on how Sunburst fits into the full release, see the related articles on the Sketch drawing feature and SynthID watermarking.
FAQ
What is GPT-Image-2.5 Sunburst?A premium OpenAI API image model released September 8, 2026, optimized for precise multi-turn editing and production-ready creative output.
How does Sunburst compare to Flare?Sunburst offers tighter iterative control and lower adversarial failure rates (1.09% vs 1.41%); Flare generates faster and targets high-volume use cases. Both cost the same.
Does Sunburst watermark images?Yes — all outputs carry C2PA metadata and an invisible SynthID watermark from Google DeepMind.