Sunburst and Flare share published rates. Choose using your own task sample: complexity and output format change what you need to validate.
What OpenAI announced this week#
On November 4, 2026, OpenAI added GPT Image 2.6 Sunburst and Flare for generation and editing. The provider positions Sunburst for detailed visual work and Flare for iteration. Those descriptions are not measured latency or quality results for your application.
Documentation: OpenAI · API changelog, 4 November 2026 ↗
How to read the cost#
These rates are USD per million tokens and are the same for both models. They are neither prices per image nor text-output rates. A request costs the sum of its actual billable token usage.
| Item | USD per million tokens |
|---|---|
| Text input | 5.00 |
| Cached text input | 1.25 |
| Image input | 8.00 |
| Cached image input | 2.00 |
| Image output | 30.00 |
Documentation: OpenAI · GPT Image 2.6 Sunburst ↗ · OpenAI · GPT Image 2.6 Flare ↗
An SVG request to start with#
{
"model": "gpt-image-2.6-flare",
"prompt": "A simple three-shape icon for a document search tool, without lettering.",
"output_format": "svg",
"n": 1,
"complexity": "medium",
"quality": "medium"
}For SVG, use n=1. The guide excludes streaming, partial images, output_compression and prompt_rewriter for this format; the edit endpoint does not produce SVG. Save the decoded XML bytes from data[0].b64_json as a .svg file.
Documentation: OpenAI · Image generation ↗
How to decide before integrating#
- Use the same brief and acceptance criteria for both models: readable lettering, composition and corrections required.
- Record billed cost, elapsed time and acceptable results per attempt; include retries when calculating cost per finished asset.
- Validate SVG in its consuming environment. Vector output does not guarantee a legible icon or compliance with your design system.
- Test raster, SVG and GIF as distinct outputs; retain an alternative format and size your interface can display.
The text LLM calculator does not estimate this image bill. Keep token types and modalities separate when comparing providers.
Evidence and scope
Documentation review · Sources reviewed:
Method
Se contrastaron el changelog del 4 de noviembre, ambas fichas de modelo y la guía de formatos de OpenAI. La tabla conserva las unidades publicadas; la solicitud y la pauta de evaluación son ejemplos de Codifly.
Limitations
- Sin llamadas pagadas ni mediciones propias de calidad o latencia. Las tarifas son por tipo de token, no por imagen. La disponibilidad depende de la cuenta.
Next review planned:
Sources and scope
Documentation checked on November 4, 2026. Examples and decision criteria are editorial proposals; adapt them to your application's contract and validate them in an authorized test environment.