GPT Image 2.6 Flare and Sunburst
When to draft on Flare and when to finish on Sunburst, with the same 20-credit cost on this site.
Official data & independent evaluations
Official specifications
Scroll horizontally to see every column ↔
| Measure | GPT Image 2.6 Flare | GPT Image 2.6 Sunburst | Source |
|---|---|---|---|
| Official positioning | Fast iteration | Detailed compositions and precise edits | OpenAI image prompting guide |
| Standard image-output token price | $30 / 1M image-output tokens | $30 / 1M image-output tokens | OpenAI API pricing |
| Standard text / image input | $5 / $8 per 1M tokens | $5 / $8 per 1M tokens | OpenAI API pricing |
| Complexity control | low / medium / high; default high | low / medium / high; default high | OpenAI image generation guide |
| GIF and SVG support | Documented for generation | Documented for generation | OpenAI image generation guide |
Arena · text-to-image
Scroll horizontally to see every column ↔
| Model / tested setting | Score ± interval | Votes / samples | Status |
|---|---|---|---|
| GPT-Image-2.6 Flare (Medium) | 1384 ±13 | 2,792 | Preliminary |
| GPT-Image-2.6 Sunburst (Medium) | 1409 ±13 | 2,967 | Preliminary |
Arena · single-image editing
Scroll horizontally to see every column ↔
| Model / tested setting | Score ± interval | Votes / samples |
|---|---|---|
| GPT-Image-2.6 Flare (Medium) | 1462 ±10 | 3,850 |
| GPT-Image-2.6 Sunburst (Medium) | 1476 ±10 | 3,833 |
Displayed snapshot: November 2, 2026; inspected November 6. Its cutoff predates the public 2.6 release, so the update alignment is unclear. Treat these as displayed historical values, not a verified November 6 ranking. Quality labels and vote counts differ.
Arena measures human preferences in head-to-head comparisons. Its scores are not percentages of image quality. Text-to-image and editing boards measure different tasks. Arena · rankings and confidence intervals
Official example: the same bakery-logo brief
OpenAI identifies both SVGs as results of the same Field & Flour logo prompt. The full prompt and matching generation settings are not published beside the examples. This illustrates two outputs, not a measured win rate.
GPT Image 2.6 Flare
GPT Image 2.6 Sunburst
In this example, Flare has a simpler loaf and lettering; Sunburst adds wheat, steam and a rolling pin. For a small logo, inspect legibility and path complexity before preferring more ornament.
What the evidence supports
Sunburst Medium has the higher displayed Arena score in both tasks. The intervals overlap, and the board is preliminary; this does not establish a precise win rate. The official pages do not provide a matched latency dataset or an editing-preservation percentage.
Use Flare to screen directions; test Sunburst when text, fine product parts or repeated edits decide acceptance. Keep prompt, references, dimensions, quality and complexity identical. Compare time to a usable file and cost per accepted output, not only the first render.