Blog · 2026-11-05

GPT Image 2.6 vs Muse Image: GIF, SVG and 4K or a Flat $0.02

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

Choose Muse Image when you want one image of about 2.5 megapixels at a flat $0.02. Choose GPT Image 2.6 when the job needs 4K, an exact pixel size, a mask, a transparent background, an animated GIF, an SVG file or several images from one call. The deciding difference is the bill: Muse charges one fixed price per delivered image, while GPT Image 2.6 is metered per token at OpenAI's official rates, so its cost moves with size, quality, format and every reference image you send.

GPT Image 2.6 is available on apimodels.app as two tiers that take the same parameters and bill from the same price list. gpt-image-2.6-flare is the default tier. gpt-image-2.6-sunburst is the tier OpenAI positions as spending more time per image on detail. In our 5 November 2026 tests, a 2K low request at 16:9 cost $0.0283 and delivered 2720×1536, and a 2K medium request cost $0.0516. Muse Image, Meta's model, returns 2048×1152 at 16:9 for $0.02, and the price is the same at every aspect ratio.

The trade is range against predictability. Muse covers still images and reference-based edits at one price. GPT Image 2.6 covers far more outputs, and you budget from measured examples instead of a fixed price.

GPT Image 2.6 vs Muse Image at a glance

Pricing — GPT Image 2.6 bills per token: image output $30 per 1M tokens, image input $8 (cached $2), text input $5 (cached $1.25) and text output $10. Each request is rounded up to $0.0001. SVG is the exception, at a flat $0.035 per image. Muse Image is a flat $0.02 per delivered image, and reference images cost nothing extra.

Resolution and sizes — GPT Image 2.6 accepts four ways to set the size: aspect_ratio with 1K, 2K or 4K; a preset name; auto; or an exact WIDTHxHEIGHT. Exact sizes must be multiples of 16, up to 3840 on the long edge, with a ratio up to 3:1 and 0.66 to 8.29 megapixels. Muse Image renders about 2.5 megapixels at a native size for each shape, such as 1600×1600 at 1:1 and 2352×1008 at 21:9. It offers auto or 13 fixed ratios. A size only sets the proportions, and anything above about 2.6 megapixels is rejected.

Inputs — GPT Image 2.6 accepts up to 16 reference images, plus either a markup image or a mask, not both. Sunburst documents prompts up to 32,000 characters. Muse Image accepts 1 to 4 reference images (PNG, JPEG or WebP, up to 10 MB each and 24 MB in total) and prompts up to 4,000 characters. It rejects masks, negative_prompt and seed.

Output types — GPT Image 2.6 returns raster images, with transparent backgrounds available. It also returns animated GIF (2 to 64 frames, 20 to 2000 ms per frame) and SVG from a text prompt, and up to 10 images per call. Muse Image returns exactly one image per request: JPEG by default, or PNG and WebP through output_format. It has no transparent background.

Speed — In our 5 November 2026 tests, GPT Image 2.6 took about 20 seconds for a 1K low text-to-image and about 27 at high quality. A 1K low edit with one reference took 22 to 58 seconds, and 2K and 4K images took 13 to 25 seconds. Seven Muse Image text-to-image requests on 30 October 2026 took 34 to 67 seconds each. We have not timed Muse edits.

Limits — GPT Image 2.6's SVG output is text-to-image only, one image per request, and quality has no effect on it. Muse Image has no seed, so results are not repeatable. It also has no quality tiers and no 4K. Results from both are kept for 30 days.

A fixed price against a token meter

Muse Image's price is one number. Each request returns one image, so $0.02 per request is also $0.02 per image. That holds for text-to-image and for an edit with four references, at 21:9 or at 1:1. Multiply it by your volume and you have the bill.

GPT Image 2.6 has no fixed price per image. Quality, size, GIF output, each reference image and the prompt all change the token count. The tier does not, because Flare and Sunburst share one price list. Our measured requests at 16:9 show the spread: about $0.0077 for 1K low, $0.0162 for 1K high, $0.0283 for 2K low and $0.0516 for 2K medium. Moving from 1K low to 2K medium added $0.0439 per image. On apimodels.app, quality defaults to medium. The model's own default is high, which uses more tokens.

Two details matter when you plan a budget. First, apimodels.app holds an upper bound when a GPT Image 2.6 request is accepted and releases the unused part when the image is ready. Your balance has to cover the hold, not only the final charge. Second, completed png, jpeg and webp requests return a usage object whose charged field is what was actually deducted. GIF, markup or mask edits and SVG come back without one.

Neither model succeeds on every call. Moderation refusals and upstream errors happen on both, and a failed request is not charged on either.

What Muse Image will not produce

For plain stills and edits, the two models overlap. Outside that, GPT Image 2.6 covers jobs Muse Image does not attempt. Muse answers those requests with a 400 before anything is created, instead of returning something smaller or silently ignoring a field, and a rejected request is not charged.

These requests are rejected by Muse Image and accepted by GPT Image 2.6:

1. A 4K output, or any size above about 2.6 megapixels, such as 3840x2160.
2. An exact pixel size. Muse treats size as proportions, so in our tests a 1024x1024 request came back at 1600×1600.
3. A mask for inpainting.
4. A transparent background.
5. n above 1. GPT Image 2.6 returns up to 10 images per call, while a Muse batch is several parallel requests.

GIF and SVG are the larger gap. GPT Image 2.6 returns an animated GIF of 2 to 64 frames, all laid out on one sheet of at most 3840 px on the long edge. With aspect_ratio plus 1K, 2K or 4K, apimodels.app shrinks the frames to fit. With an exact size, you keep it small yourself; the whole sheet must stay within 3840 px on the longest edge and 8,294,400 pixels in total. SVG output costs a flat $0.035 and is a static vector with no scripts, external resources or embedded raster. It suits icons, logos and flat illustrations rather than photographic scenes. Muse has no equivalent for either.

Editing: four references or sixteen

Both models edit through the same request as text-to-image. Muse Image takes 1 to 4 references, and the first one is the image to change. You point at the others in the prompt, for example by asking it to put the cup from the second image on the table. Every edit costs $0.02. There is no mask, so you describe the change in words.

The Muse model page adds two cautions. Small text inside a reference image may not be preserved. Long passages of text in the output can contain typos, so add exact long copy afterwards. We have not run our own Muse image-to-image tests, so we have no measured editing results to quote.

GPT Image 2.6 takes up to 16 references and gives you two ways to mark a region. A markup image is a transparent PNG with circles, arrows or notes drawn over the first reference. A mask uses white for the area to change and black for the area to keep. You cannot send both. On the previous generation, masks behaved as hints rather than hard boundaries, and we have not measured 2.6 yet, so name the region in the prompt as well. References are billed as image input tokens. Our 1K low edit with one reference cost about $0.0157.

For edits that run over many rounds, Sunburst is built to keep edits scoped to the instruction, with subject and composition preserved across revisions. It bills from the same price list as Flare. The difference is time per image, not price.

What a job costs

The Muse figures below are fixed. The GPT Image 2.6 figures multiply our measured single requests from 5 November 2026 at 16:9. They are examples, not a fixed price, and each real charge depends on the tokens that request uses.

100 product shots at 16:9

Muse Image returns 100 images at 2048×1152 for $2.00. GPT Image 2.6 comes to about $0.77 for 100 images at 1K low (1360×768), about $2.83 at 2K low (2720×1536) and about $5.16 at 2K medium. Per image, 2K low costs $0.0083 more than Muse and delivers more pixels. 1K low costs $0.0123 less and delivers fewer.

100 edits with one reference image

Muse Image costs $2.00, since references are free. GPT Image 2.6 at 1K low comes to about $1.57, based on our measured single-reference edit. Muse renders each edit at about 2.5 megapixels, while the GPT figure is at 1K. A larger size or more references will raise the GPT figure.

10 SVG icons

GPT Image 2.6 costs $0.35: one request per icon at the flat $0.035, whatever the quality. Muse Image cannot return SVG or a transparent background, so it is not a substitute here.

One 2K hero image, inspected up close

Sunburst at 2K medium measured $0.0516, which is 2.58 times Muse's $0.02 for a 2048×1152 image. If the hero needs 4K, only GPT Image 2.6 offers it. 4K uses more tokens than 2K, so expect a 4K image to cost more than the 2K figure. We have not published a measured 4K or GIF cost.

Which one to use

Pick Muse Image when you produce still images in volume, about 2.5 megapixels is enough, and you want to know the cost of a batch before it runs. It also fits edits that combine up to four pictures by description, at the same $0.02. The trade-offs are that results are not repeatable without a seed and that each request returns only one image.

Pick GPT Image 2.6 Flare when you need 4K or an exact pixel size, transparent backgrounds, masks or markup edits, more than four references, animated GIF, SVG, or up to 10 images per call. In our measurements it was also the lower-priced option for 1K drafts. Pick Sunburst for packaging, print, product detail and dense type, or for edits that go through several rounds. The price list is identical, and the cost follows the quality and size you request.

Do not move a GPT Image 2.6 pipeline that already runs at 1K low over to Muse to save money. Our measured $0.0077 per image costs less than Muse's $0.02. If you want OpenAI's model at a fixed per-image price, neither of these is the direct answer. GPT Image 2.5 stays available on apimodels.app with fixed per-image prices, and it is the one to check when you do not need the 2.6 additions.

Calling both with one key

One apimodels.app key and one endpoint cover all three model ids. Only the model field and the options change.

# GPT Image 2.6 Flare: an animated GIF, 9 frames at 200 ms each
curl -X POST 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": "a paper boat rocking on small waves, flat colours", "aspect_ratio": "16:9", "output_format": "gif", "frame_count": 9, "frame_duration_ms": 200}'
# GPT Image 2.6 Sunburst: high quality for a detailed product shot
curl -X POST 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": "a glass perfume bottle on brushed steel, studio light", "aspect_ratio": "16:9", "quality": "high"}'
# Muse Image: edit one reference image at the flat $0.02
curl -X POST https://api.apimodels.app/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "muse-image", "prompt": "keep the person and their position, change the background to a beach, oil-painting style", "images": ["https://example.com/portrait.jpg"], "aspect_ratio": "16:9"}'
# Each call returns a taskId: poll GET /v1/images/generations?task_id=... until state is completed

FAQ