GPT Image 2.6 API Tutorial: Flare vs Sunburst, GIF and SVG Output

Flare or Sunburst: Pick Your Model ID First

Start with the model ID. GPT Image 2.6 ships as four IDs, and the choice comes down to speed versus edit consistency — not price. Flare and Sunburst share identical parameters, usage patterns, and pricing. The difference is temperament.

• gpt-image-2.6-flare/text-to-image — fastest path from prompt to pixels. Default choice for prompt iteration, A/B drafts, and batch jobs.

• gpt-image-2.6-flare/image-to-image — same speed, with source images attached.

• gpt-image-2.6-sunburst/text-to-image — heavier detail model. Reach for it when the first frame matters more than the tenth draft.

• gpt-image-2.6-sunburst/image-to-image — the one to use for multi-round edits where characters, products, or layouts must stay consistent across passes.

A simple rule: if you are still deciding what the image should look like, use Flare. Once you are refining a look across several edits at medium quality or above, switch to Sunburst. Because parameters and pricing are identical, moving between them is a one-line change in your request body — no schema rewrite.

{
  "model": "gpt-image-2.6-flare/text-to-image",
  "prompt": "a neon-lit ramen stall in the rain, cinematic"
}

Swap the model string and you are on Sunburst. That is the whole migration.

Text-to-Image Parameters That Change the Output

Most parameters are independent axes, not a single dial. The three that matter most are quality, complexity, and resolution.

quality

low, medium, high, xhigh, max. This drives price and detail. Default sensibly: low or medium during iteration, high and above when the frame is final.

complexity

low, medium, high — a planning-level control, independent of quality. Default is high. It does not substitute for quality.

resolution and aspect_ratio

resolution is 1K, 2K, or 4K (default 1K). aspect_ratio accepts auto or any of 15 explicit ratios. Together they shape the canvas, not the rendering effort.

output_format and output_compression

output_format is png (default), jpeg, webp, gif, or svg. output_compression runs 0–100 and applies only to jpeg and webp. Setting it on PNG does nothing — do not ship dead parameters.

background

auto, opaque, or transparent. transparent requires png or webp; don't pair it with jpeg, gif or svg.

A minimal, deliberate set:

{
  "model": "gpt-image-2.6-flare/text-to-image",
  "prompt": "isometric coffee roastery, soft morning light",
  "quality": "medium",
  "resolution": "2K",
  "aspect_ratio": "16:9",
  "output_format": "png",
  "background": "opaque"
}

Every parameter you omit takes a sane default. Every parameter you set should earn its place.

Image-to-Image: Source Images, mask_url and markup_url

The image-to-image IDs take 1–16 source images in jpeg, png, or webp, passed as image_urls. That is the input contract — no more, no fewer.

On top of that you get exactly one editing hint: mask_url or markup_url, never both.

mask_url

An RGBA PNG at the same dimensions as the first source image. Transparent pixels mark the region to change. Use it when you know where to edit pixel-precisely.

markup_url

A transparent PNG where you draw annotations circling the area to edit. The annotation guides the model and does not appear in the result. Use it when you want to describe a region loosely instead of cutting a precise mask.

Two rules to keep in mind:

• The two fields are mutually exclusive. Sending both is a validation error.

• Masking is a hint, not a lock. Areas outside the mask can still shift slightly. Do not promise pixel-perfect preservation downstream.

{
  "model": "gpt-image-2.6-sunburst/image-to-image",
  "image_urls": ["https://example.com/product.png"],
  "mask_url": "https://example.com/product-mask.png",
  "prompt": "replace the background with matte concrete",
  "quality": "high"
}

Output is limited to png, jpeg, or webp — no gif or svg on this path.

Configuring GIF Output: frame_size, frame_count, frame_duration_ms

GIF is a different beast: it takes its own parameters and has hard limits enforced at submit time.

frame_size

Width-by-height in pixels, e.g. 512x512. Each side must be a multiple of 16. Default is 512x512. It also defines the GIF's aspect ratio — so do not send aspect_ratio separately; leave it unset or set it to auto. The human-readable 1:1 style is not accepted here; use WIDTHxHEIGHT.

frame_count

From 2 to 16 frames. Default is 9. At 768x768, the cap drops to 12 frames.

frame_duration_ms

Minimum 20, default 200. Values are rounded to 10 ms automatically.

The hard ceilings

• Each side of the whole stitched frame sheet ≤ 3840 (the sheet, not a single frame).

• Total pixels ≤ 8,294,400.

• Aspect ratio ≤ 3:1.

Exceed any of these and the request is rejected at the submit stage — no image, no charge. Validate client-side before sending: check the multiple-of-16 rule, the frame-count range, and the sheet's side, total-pixel and ratio caps. For reference, 16 frames at 512x512 stitch into a 4×4 sheet of 2048×2048 (about 4.19 million pixels), well inside the limits; 768x768 tops out at 12 frames.

GIF is billed by quality only: low $0.070, medium $0.124, high $0.150, xhigh $0.266, max $0.598. The example below costs $0.124.

{
  "model": "gpt-image-2.6-flare/text-to-image",
  "prompt": "pixel-art fox walking through snow, seamless loop",
  "output_format": "gif",
  "frame_size": "512x512",
  "frame_count": 12,
  "frame_duration_ms": 120,
  "quality": "medium"
}

Note: output_format: "gif" is text-to-image only.

Requesting SVG Output and What It Costs

SVG output is text-to-image only. Set output_format to svg on a text-to-image ID and nothing else changes in the request shape.

{
  "model": "gpt-image-2.6-sunburst/text-to-image",
  "prompt": "minimal line icon of a paper plane, single color",
  "output_format": "svg",
  "quality": "medium",
  "resolution": "1K"
}

The cost model

SVG is billed as the same quality-tier PNG price plus a flat $0.030. It is a surcharge on the tier, not a separate price table: quality and resolution set the base, and the $0.030 is the same on every tier.

So the example above (medium 1K) costs $0.041 + $0.030 = $0.071. At the extremes, low 1K SVG is $0.053 and max 4K SVG is $0.666. The full table is on the pricing page.

The practical implication: use low or medium for icon and logo drafts, and reserve high and above for final vector assets where the detail has to survive scaling.

Cost Planning: Per-Image Pricing and Refunded Rejections

Pricing rules to internalize before you build a budget:

• Standard image output is billed per image by quality and resolution. There is no per-token accounting to reason about — pick the tier, get the price:

• image-to-image charges the same table plus $0.012 for each source image you attach. Attaching 16 images adds $0.192.

• GIF is billed by quality alone, regardless of resolution: low $0.070, medium $0.124, high $0.150, xhigh $0.266, max $0.598.

• SVG is the same-tier PNG price plus $0.030.

• Rejected submissions — invalid parameters, GIF dimension overruns — are fully refunded. You pay for successful renders, not attempts.

That last point changes how you should think about validation. Client-side checks cost you latency and nothing else; a malformed request costs the same as a round trip. Build the validator and skip the guessing.

When comparing against the older generation, note that GPT Image 2.5 standard lines are billed by resolution only (1K $0.025, 2K $0.040, 4K $0.060) rather than the quality-and-resolution grid used here.

A Pre-Flight Checklist Before You Ship

Run this list against every request path before your first production call.

• Model ID: Flare for iteration and batch; Sunburst for consistent multi-round edits at medium or above.

• Format and path: gif and svg only on text-to-image. Image-to-image outputs png, jpeg, or webp — nothing else.

• Compression: output_compression only with jpeg or webp. Strip it from PNG requests.

• Transparency: background: "transparent" requires png or webp — never with gif, svg or jpeg.

• GIF geometry: frame_size sides are multiples of 16 and written as WIDTHxHEIGHT; frame_count is 2–16 (12 max at 768x768); frame_duration_ms ≥ 20, rounded to 10 ms automatically. No aspect_ratio alongside GIF.

• GIF ceilings: the stitched frame sheet stays ≤ 3840 per side, ≤ 8,294,400 total pixels, ratio ≤ 3:1. These reject at submit.

• Masks: mask_url and markup_url are mutually exclusive. mask_url is an RGBA PNG matching the first source image.

• Source images: 1–16 files in image_urls, jpeg / png / webp.

• Budget: GIF priced by quality only ($0.070–$0.598); SVG adds $0.030; image-to-image adds $0.012 per reference image; rejects are refunded.

The fastest way to verify a single path is to open the model page for Flare text-to-image and run one call with defaults. Ship the checklist as a unit test, and the launch is quiet.