GPT Image 2.6 is OpenAI’s November 4 image-generation and editing release. Developers choose Sunburst for detailed visual work or Flare for fast iteration. Both have dedicated model IDs, and the current model cards list the Image API rather than Chat Completions or Responses as their supported integration surface.
This guide covers the developer workflow: choosing a model, budgeting, controlling edits and handling outputs. It is based on the official image-generation documentation, not an independent quality or latency benchmark. ChatGPT and Codex product interfaces can expose different controls from the public Image API.
Model IDs and endpoints
| Property | Sunburst | Flare |
|---|---|---|
| Model ID | gpt-image-2.6-sunburst | gpt-image-2.6-flare |
| Documented emphasis | Detailed visual work | Faster iteration |
| Input | Text and image | Text and image |
| Output | Image | Image |
| Generation | /v1/images/generations | /v1/images/generations |
| Editing | /v1/images/edits | /v1/images/edits |
| Responses / Chat Completions | Not listed as supported | Not listed as supported |
Use the Sunburst card and Flare card to confirm endpoint support. Do not assume that because an older image model appears as a tool inside another endpoint, these new IDs work there too.
Keep the exact model name in configuration. Test generation and editing separately, including error handling for an unsupported format or parameter. A valid image request does not prove that your client serializes every new editing option correctly.
Pricing: text and image tokens are different categories
The model cards currently show the same token prices for Sunburst and Flare:
| Category | USD per million tokens |
|---|---|
| Text input | $5 |
| Cached text input | $1.25 |
| Image input | $8 |
| Cached image input | $2 |
| Image output | $30 |
These are token rates, not a flat price per image. Resolution, quality, reference images and output format can change the amount of work and billed usage. Do not multiply a text prompt’s token count by the image-output price and call that the full generation cost.
For a real budget, record usage from representative accepted assets. Include discarded candidates and revision rounds. A cheaper first attempt can be more expensive if it takes several retries to meet the requirements. The image API pricing guide covers the cross-provider accounting problem; Get API Pulse keeps structured media-price records separate from ordinary text-model comparisons.
Complexity and quality control different work
The 2.6 complexity setting accepts low, medium and high, with high as the documented default. It controls planning effort. quality controls rendering detail. Treat them as separate variables in a trial.
For a narrow edit, begin by evaluating low complexity at the quality your output needs. For a scene with many interacting constraints, compare higher complexity against the same prompt. Neither setting is a guarantee that every constraint will be followed.
Record both settings alongside the output so a later reviewer can reproduce the intended configuration. Do not describe a quality change as a model upgrade, or compare one model’s low-effort attempt with another’s most expensive configuration without disclosing the difference.
Editing and markup guidance
The image guide documents markup-guided edits for supported models. An annotation can indicate the region or object a reviewer means, while the prompt describes the desired result. That is useful when replacing a label, changing a product detail or refining a composition.
An annotation is guidance, not proof that unmarked pixels remain identical. Review the whole resulting image for changed text, geometry, brand details and faces. Keep the original and the edit together during approval. An attractive local change can still damage another part of the asset.
Build the workflow around explicit acceptance criteria: required text, fixed product details, allowable edits and intended destination. Our AI testing foundation explains why a visible result and a model’s confident description are different evidence.
SVG output needs a separate acceptance path
The Image API supports SVG output for the documented 2.6 models. An SVG is a structured asset that may need review beyond checking a raster preview. Confirm that it renders correctly in the browsers or design tools where it will be used.
Treat generated SVG as untrusted active content before embedding it directly in a website. Use an appropriate sanitization and serving policy, inspect external references and check accessibility. The fact that a model generated the file does not make every element suitable for your application.
Also verify the actual editing requirement. A file that opens as SVG is not automatically a neatly organized illustration with useful layers, a small path count or a reusable logo system. Check the output against the downstream job rather than promising editability from a format name alone.
Animated GIF is not a general video model
The documentation also supports animated GIF output. Test motion continuity, loop behavior, dimensions, palette and file size in the actual destination. A short animation that looks acceptable locally may exceed a messaging or publishing limit.
Do not infer long-form video generation, synchronized audio or an unrestricted video timeline from GIF support. Those are separate product capabilities. Preserve a static fallback and account for accessibility preferences when publishing motion.
For a product pipeline, compare accepted animations rather than the best handpicked sample. Evaluate character consistency over frames and repeated requests. The regression testing guide provides a way to retain difficult examples instead of relying on memory of a good demonstration.
Access, privacy and reliability
API credentials and ChatGPT subscription access are separate arrangements. Check your organization’s available models, limits and billing before sending requests. The model cards state that rate limits depend on account and usage tier; there is no universal throughput promise here.
Reference images can contain personal data, customer material or confidential product designs. Apply your organization’s upload, retention and access policy. Do not equate image editing with local processing. See the data-retention guide and agent security guide for controls around applications that generate or publish assets.
My take
Try Flare and Sunburst on the same real generation and edit tasks. Keep complexity, quality and acceptance criteria explicit, then compare cost per approved asset. The most useful additions are controllable iteration and new output formats; their value depends on whether the resulting files survive your production checks.
FAQ
Is Sunburst more expensive per token than Flare?
The checked model cards show identical token rates. Actual request cost can differ with usage and configuration.
Can I call these model IDs through Chat Completions?
The current cards list generation and editing through the Image API, not Chat Completions or Responses.
Does markup preserve every unmarked pixel?
It is guidance for the edit. Inspect the complete result; exact preservation is not established by an annotation.
Does GIF support mean full video generation?
No. Evaluate the documented animation workflow without inferring audio or general video capabilities.