Blog · 2026-11-05

GPT Image 2.6 vs GPT Image 2.5: New Outputs or a Fixed Image Price

Model pages: GPT Image 2.6 Flare API · GPT Image 2.6 Sunburst API · GPT Image 2.5 Flare API · GPT Image 2.5 Sunburst API — pricing, specs and the playground

GPT Image 2.6 is the model to call when the job needs animated GIF, SVG, markup-image or mask edits, or an exact pixel size; GPT Image 2.5 is the one to call when you want to know an image's price before you send the request. The deciding fact is billing: 2.6 is charged per token, with image output at $30 per 1M tokens, while 2.5 charges a fixed cell on a resolution × quality grid starting at $0.01.

So which one costs less depends on size, not generation. In our tests on 5 November 2026, a 1K low request on 2.6 came to about $0.0077, below the $0.01 that 2.5 charges at 1K low, while a 2K medium request came to $0.0516 against a fixed $0.03 on 2.5. 2.6 is a new set of abilities rather than a sharper 2.5, and 2.5 keeps quality levels that 2.6 lacks.

GPT Image 2.6 vs GPT Image 2.5 at a glance

Tiers — Each generation has two tiers, Flare as the default and Sunburst for the highest fidelity. In 2.6 they take the same parameters and bill from the same price list, so cost follows what you ask for, not the tier. In 2.5 they share every cell at 1K and 2K, but at 4K Sunburst runs $0.02 / $0.085 / $0.16 / $0.28 / $0.47 from low to max against $0.02 / $0.045 / $0.08 / $0.17 / $0.3 on Flare.

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 rounded up to $0.0001; SVG is a flat $0.035 per image. GPT Image 2.5 charges a fixed price per image from fifteen resolution × quality cells, $0.01 to $0.3 on Flare and $0.01 to $0.47 on Sunburst, with reference images and transparent backgrounds included.

Resolution and sizes — 2.6 accepts 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. 2.5 renders native 1K, 2K and 4K in six aspect ratios (1:1, 3:4, 4:3, 9:16, 16:9, 21:9) and bills by the pixel count of the returned image: 1.7MP or less is 1K, 1.7-4.6MP is 2K, above 4.6MP is 4K.

Quality — 2.6 offers low, medium, high or auto, plus a separate reasoning_effort (low, medium or high) that trades latency for detail; both 2.6 tiers default to medium on apimodels.app. 2.5 offers low, medium, high, xhigh and max as a price axis, defaulting to medium on Flare and high on Sunburst.

Inputs — 2.6 takes up to 16 reference images plus either a markup image or a mask, not both. 2.5 takes reference images, but a request carrying a mask returns 400, so local edits are steered through the prompt.

Output types — 2.6 returns PNG, JPEG or WebP with transparent backgrounds available, plus animated GIF and SVG, up to 10 images per call. 2.5 returns raster images, including native transparent PNG; GIF and SVG are not part of it.

Speed — In our 5 November 2026 tests, 2.6 took about 20 seconds at 1K low, about 27 seconds at 1K high, 13 to 25 seconds at 2K and 4K, and 22 to 58 seconds for a 1K low edit with one reference. Single 2.5 test calls took 37s at 1K low, 17s at 2K medium and 95s at 4K high on Flare, and 36s at 1K high, 17s at 2K high and 28s at 4K high on Sunburst, the last on a previous channel.

Limits — On 2.6, SVG is text-to-image only, GIF frames share one sheet of at most 3840 px on the longest edge, and results are kept for 30 days. On 2.5, sketch input, the creation templates and comment-on-image editing are ChatGPT features, not part of the API.

Per-token billing against a fixed grid

With 2.5 you can price a job before running it: pick a size and a quality, read the cell, multiply. A 2K medium banner is $0.03 on either tier and a 4K high poster is $0.08 on Flare or $0.16 on Sunburst. The tier is set by the pixel count of the image you get back, not by the parameter you sent, so a 2K image at 2720x1536 is 4.18MP and bills as 2K.

With 2.6 the price is assembled from tokens. Quality, size, GIF, every reference image and the prompt all change the count, so there is no fixed price per image. On apimodels.app an upper bound is held when the request is accepted and the unused part is released when the image is ready. Completed PNG, JPEG and WebP requests return a usage object whose charged field is what was deducted; GIF, markup or mask edits, SVG and failed tasks return none.

Our 16:9, one-image requests on 5 November 2026 came to about $0.0077 at 1K low, $0.0162 at 1K high, $0.0157 for a 1K low edit with one reference, $0.0283 at 2K low and $0.0516 at 2K medium. These are examples from individual requests, not prices. For text-to-image, 2.6 came in below the 2.5 grid at 1K and above it at 2K. No 4K cost for 2.6 is documented, so a 4K comparison needs your own test call.

What 2.6 adds that 2.5 does not have

The case for 2.6 is a set of abilities that 2.5 does not offer:

1. Animated GIF: output_format gif, with frame_count from 2 to 64 (default 9) and frame_duration_ms from 20 to 2000 (default 200).
2. SVG: output_format svg returns a static vector from a text prompt at a flat $0.035, suited to icons, logos and flat illustrations.
3. Markup-image edits: circles, arrows or notes drawn on a transparent PNG show the model what to change in the first reference.
4. Masks: white marks the area to change, black the area to keep; 2.5 rejects any mask.
5. Batch and effort controls: up to 10 images per call, up to 16 references, and reasoning_effort (low, medium or high).

Each comes with a constraint. GIF frames share one sheet of at most 3840 px on the longest edge; pass aspect_ratio with 1K, 2K or 4K, or no size, and apimodels.app shrinks the frames to fit, while 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 is one image per request, quality has no effect on it, and the file holds no scripts, external resources or embedded raster. A markup image cannot be combined with a mask.

In our tests of the previous generation, masks behaved as hints rather than hard boundaries, which is why 2.5 returns 400 for them. We have not measured how strictly 2.6 respects a mask, so for a strict local edit also name the region in the prompt, for example "change only the label on the bottle, keep everything else unchanged".

Where 2.5 still holds its ground

The xhigh and max quality levels exist only on 2.5: 1K max is $0.08, 2K max is $0.15, and 4K max is $0.3 on Flare and $0.47 on Sunburst. 2.6 stops at high and offers reasoning_effort as a separate dial, so a workflow built on max quality has no direct equivalent there.

Local edits work without a mask. Name the object to change and say the rest stays; in our tests, "change only the red circle to blue, leave everything else unchanged" changed only the circle. Native transparent PNG, with background set to transparent and output_format to png, is verified at 1K, 2K and 4K at no extra charge.

Watch the defaults when you switch: 2.5 Sunburst renders at high unless told otherwise, while 2.6 Sunburst defaults to medium on apimodels.app, so pass quality explicitly.

What a job costs

The 2.5 figures are fixed cell prices for either tier. The 2.6 figures multiply our measured examples from 5 November 2026, so treat them as estimates; prompt length and references will move them.

1. 100 social images at 1K low, 16:9: about $0.77 on 2.6, $1 on 2.5.
2. The same 100 at 1K high: about $1.62 on 2.6, $3 on 2.5.
3. 100 single-reference edits at 1K low: about $1.57 on 2.6, $1 on 2.5, reference included.
4. 100 category banners at 2K medium, 16:9: about $5.16 on 2.6, $3 on 2.5; at 2K low, about $2.83 against $2.
5. A set of 10 SVG icons: $0.35 on 2.6; 2.5 has no SVG output.

2.6 has no measured 4K or GIF cost yet, so run a test call before budgeting those. On either generation a request can fail through a moderation refusal or an upstream error, and failed requests are not charged.

Which one to use

Pick GPT Image 2.6 if the deliverable is a GIF or an SVG, if you want to point at the change with a markup image or a mask, if you need an exact pixel size, or if your text-to-image volume sits at 1K, where our measured requests came in below the 2.5 grid. Use Flare for everyday volume and Sunburst when someone will zoom in; the per-token price is the same.

Pick GPT Image 2.5 if you need the price before the call, if your volume sits at 2K, if you rely on xhigh or max quality, or if you render 4K and want a documented per-image price. Flare is the natural home for 4K volume, since Sunburst's 4K row is higher.

Do not move a working 2.5 pipeline at 2K medium just because 2.6 is newer: in our tests 2.6 billed $0.0516 for a size that costs $0.03 on 2.5, and you gain nothing unless you need one of the new outputs or controls. If the job is turning an existing logo or photograph into a vector, neither fits: 2.6 makes SVG from a text prompt only and suits flat artwork rather than photographic scenes, and 2.5 has no SVG output.

Calling both with one key

All four models share one endpoint and one apimodels.app key; only the model value and the request fields change.

# GPT Image 2.6 Flare: animated GIF, 9 frames at 200 ms
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": "A paper boat on small waves", "output_format": "gif", "frame_count": 9, "frame_duration_ms": 200, "size": "960x960"}'

# GPT Image 2.6 Sunburst: 2K, high quality, reasoning effort high
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": "Wool coat with brass buttons", "aspect_ratio": "16:9", "size": "2K", "quality": "high", "reasoning_effort": "high"}'

# GPT Image 2.5 Flare: transparent PNG at 1K low ($0.01)
curl https://api.apimodels.app/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-image-2.5-flare", "prompt": "A ceramic mug, product shot", "aspect_ratio": "1:1", "size": "1K", "quality": "low", "background": "transparent", "output_format": "png"}'

# GPT Image 2.5 Sunburst: 4K high ($0.16)
curl https://api.apimodels.app/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-image-2.5-sunburst", "prompt": "Trade-show poster, engraved brass", "aspect_ratio": "16:9", "size": "4K", "quality": "high"}'

# Each call returns a taskId; poll it with
# GET https://api.apimodels.app/v1/images/generations?task_id=YOUR_TASK_ID

FAQ