Blog · 2026-11-05

GPT Image 2.6 Flare vs Sunburst: Same Price, Volume vs Detail

Model pages: GPT Image 2.6 Flare API · GPT Image 2.6 Sunburst API — pricing, specs and the playground

GPT Image 2.6 Flare and GPT Image 2.6 Sunburst take the same parameters and bill from the same per-token price list, so choosing between them is about how much time and detail each image deserves, not about cost. Pick Flare, the default 2.6 tier, for social posts, catalogue images and high-volume pipelines. Pick Sunburst, the highest-fidelity tier, for premium visuals that someone will inspect up close and for edits that must stay scoped to the instruction across several rounds.

The tier has no price of its own, and that decides most cases. On apimodels.app both models charge $30 per 1M image output tokens, $8 per 1M image input tokens, and $5 in and $10 out per 1M text tokens. In our 5 November 2026 tests a 1K low image came to about $0.0077 and a 2K medium image to $0.0516, and those figures are documented for both tiers. Quality, size, format and reference images move the bill; the model name does not.

One caveat before you plan around speed. OpenAI positions Sunburst as the tier that spends more time per image on detail, but the timings we documented for the two tiers fall in the same ranges, so there is no measured per-tier latency gap to report yet.

GPT Image 2.6 Flare vs Sunburst at a glance

Pricing — Flare bills per token: image output $30 / 1M, image input $8 (cached $2), text input $5 (cached $1.25), text output $10, each request rounded up to $0.0001; SVG is a flat $0.035 per image. Sunburst uses the identical list. On both, an upper bound is held when the request is accepted and the unused part is released when the image is ready.

Positioning — Flare is the default tier, described as fast, high-quality generation with natural lighting, rich textures and complex layouts, and the tier to build a pipeline on. Sunburst is the precision tier: more time per image for fine detail such as fabric weave, fine type and reflective surfaces, with edits that keep subject and composition across many rounds of revision.

Resolution and sizes — Both accept aspect_ratio with 1K, 2K or 4K, an exact WIDTHxHEIGHT (multiples of 16, longest edge up to 3840, ratio up to 3:1, 0.66 to 8.29 megapixels), a preset name or auto. In our tests a 16:9 request came back at 1360×768 at 1K on both tiers, and at 2720×1536 at 2K on Sunburst.

Inputs — Both take up to 16 reference images plus either a markup image, a transparent PNG with circles, arrows or notes drawn over the first reference, or a mask where white marks the area to change and black the area to keep. The two cannot be combined.

Outputs — Both return png, jpeg or webp, animated GIF (2 to 64 frames, 20 to 2000 ms per frame) and SVG from a text prompt, with up to 10 images per call.

Controls — Quality is low, medium, high or auto, defaulting to medium on apimodels.app; the model's own default is high, which uses more tokens. reasoning_effort takes low, medium or high and trades latency for detail.

Speed — The same figures are documented for both tiers from our 5 November 2026 tests: about 20 seconds for a 1K low text-to-image, about 27 seconds at 1K high, 22 to 58 seconds for a 1K low edit with one reference, and 13 to 25 seconds for 2K and 4K, high quality included.

Limits — SVG is text-to-image only, one image per request (n must be 1), and quality has no effect on it. GIF frames share one sheet of at most 3840 px on the longest edge. Results are kept for 30 days.

The tier is not a cost lever

Because Flare and Sunburst read from one price list, moving a workload between them does not change the rates you pay. The cost follows what you ask for, not the tier, which makes the settings inside the request the real cost controls.

The measured examples show how far those settings move a request. At 1K, raising quality from low to high took a request from about $0.0077 to about $0.0162, roughly twice as much. At 2K, medium cost $0.0516 against $0.0283 for low, a little under twice. Turning a 1K low request into an edit with one reference image brought it to about $0.0157, around $0.008 more, because the reference is metered as image input at $8 per 1M tokens.

So if a Sunburst bill looks higher than a Flare bill, check quality, size and references before blaming the tier. apimodels.app defaults quality to medium; code that passes high, the model's own default, uses more tokens on either tier.

You can confirm the figure per request. Completed png, jpeg and webp text-to-image and edit requests return a usage object in the OpenAI Images shape, in data.usage next to credits on the task query. Its cost block has amount, discount and charged; charged is what was actually deducted and is the number to rely on. GIF, markup or mask edits, SVG and failed tasks return no usage.

Where Sunburst earns its place: close-up detail and repeated edits

The documented difference between the tiers is what each is built for. Sunburst is positioned for work that will be looked at closely, such as packaging, print, product detail and dense type, and for edit discipline: changes held tightly to the instruction, with subject and composition preserved over many rounds. If your workflow is a designer requesting revision after revision on one product shot, that second property is the one to test.

No side-by-side measurement of detail between the two tiers is documented, so treat this as positioning to verify on your own images. Run the same prompt at the same quality and size on both models and compare at full resolution; the token rates are the same on either side.

For local edits, both tiers give you two ways to point at the change: a markup image drawn over the first reference, or a mask with white where the model should change the image and black where it should keep it. On the previous generation masks behaved as hints rather than hard boundaries, and mask strictness on 2.6 has not been measured yet, so name the region in the prompt as well, for example asking to change only the label on the bottle and keep everything else unchanged.

For high-volume work none of this is needed. Flare is the documented default for social posts, catalogue images and pipelines, and it carries every 2.6 ability Sunburst has.

Speed: what was measured and what was not

The timings listed above are the same for both tiers. In those runs 2K and 4K at high quality came back in 13 to 25 seconds, against about 27 seconds for 1K high, so do not assume latency grows with resolution. The single-reference edit varied most, from 22 to 58 seconds.

OpenAI's description says Sunburst spends more time per image; our documented numbers do not show that gap. If your pipeline has tight timeouts or a user waiting on screen, measure Sunburst with your own prompts and settings before switching. On both tiers, reasoning_effort (low, medium or high) is the documented lever for trading latency against detail.

Generation can fail on either tier. Moderation refusals and upstream errors happen, and failed requests are never charged on apimodels.app.

What a job costs

These are measured examples from our 5 November 2026 tests (16:9, one image per request), not fixed prices. Prompt length and references change the token count, and the examples apply to both tiers because the price list is shared.

1. 100 social images at 1K low: about $0.77 on Flare or Sunburst.
2. 100 product shots at 1K high: about $1.62 on either tier.
3. 10 hero images at 2K medium: about $0.516, or about $0.283 at 2K low.
4. 100 single-reference edits at 1K low: about $1.57.

SVG is the one fixed price: 100 icons from text prompts cost $3.50 on either tier at a flat $0.035 each, whatever the quality, sent one per request. GIF has no documented example cost and returns no usage object, so test a short animation before committing a batch.

Which one to use

Pick Flare when images are produced in volume and seen at feed or catalogue size: social posts, listing images, drafts, variants, and pipelines where steady throughput matters. It is the documented default tier and has the full 2.6 feature set, including GIF, SVG, markup edits and up to 10 outputs per call.

Pick Sunburst when the image is the product: packaging, print, close-up product detail, dense type, or a hero visual that will be zoomed into. It is also the tier to test when an edit must touch only what you asked for across several rounds. Pass high quality when detail is the point; that setting, not the tier, is what raises the token count.

Neither tier is right if you need a predictable price per image for quotes or budgets: 2.6 has no fixed per-image price outside SVG, while GPT Image 2.5 stays available on apimodels.app with fixed per-image prices. Neither suits photographic vectors either, because SVG from both tiers is a static file with no embedded raster, meant for icons, logos and flat illustrations. And if you already run Flare and want to lower spend, do not switch tiers; lower quality or size instead.

Calling both with one key

Both tiers sit behind the same apimodels.app key and OpenAI-compatible base URL, so switching is a change to the model string. The two requests below differ only in model, prompt and quality.

# GPT Image 2.6 Flare: everyday volume at the default medium quality
curl https://api.apimodels.app/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.6-flare",
    "prompt": "Ceramic mug on a linen tablecloth, soft morning light, catalogue style",
    "aspect_ratio": "16:9",
    "quality": "medium"
  }'

# GPT Image 2.6 Sunburst: same fields, high quality for a close-up hero image
curl https://api.apimodels.app/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.6-sunburst",
    "prompt": "Close-up of a perfume bottle label with fine serif type on brushed metal",
    "aspect_ratio": "16:9",
    "quality": "high"
  }'

# Each call returns a taskId; poll GET /v1/images/generations?task_id=TASK_ID
# until the image is ready. data.usage on the task query shows what was charged.

FAQ