Skip to content

GPT-image 2.6 โ€‹

Updated

โ† Back to the GPT-image overview for a version comparison and selection guide.

GPT-image 2.6 is the latest generation of the GPT-image family and ships two variants on TokenHub:

Model IDPositioningBest for
custom-model-og-v2.6-fFlare โ€” fasterFast iteration, batch drafts, interactive previews. As fast as roughly 6 seconds per image at the low tier
custom-model-og-v2.6-sSunburst โ€” more detailComplex compositions, fine-grained editing, keeping characters and products consistent across multiple rounds of revision

The two share identical request fields and the same token unit price. Compared with 2.5, overall generation is about 50% faster.

This page is one of three GPT-image API documents. See also GPT-image 2.0 and GPT-image 2.5.

Prerequisites โ€‹

  1. You have registered a Tencent Cloud account and activated the TokenHub service.
  2. Your account has been allowlisted for this model.
  3. You have obtained an API Key. How to do it: go to Online Inference โ€“ Vision Models, search for custom-model-og-v2.6-f and custom-model-og-v2.6-s, enable pay-as-you-go, then create a Key for the model under API-Key management.

Important

  • This model is a test model and requires an allowlist entry granted by the backend team before it can be used. Contact the product team and submit your UIN and AppId.
  • Allowlisting takes effect per model ID โ€” -f and -s must be requested separately.
  • GIF output requires an additional GIF preview entitlement for the model.
  • TokenHub settles on an hourly cycle. Sales should apply for customer discounts in good time.

Replace YOUR_API_KEY in every example below with your own API Key. Authentication uses the request header Authorization: Bearer YOUR_API_KEY.

Basic Information โ€‹

ItemValue
modelcustom-model-og-v2.6-f / custom-model-og-v2.6-s
Supported tasksText-to-image, image editing, markup-based editing, animated GIF, SVG vector
Request domain (international site)tokenhub-intl.tencentcloudmaas.com (Singapore)
AuthenticationAuthorization: Bearer <API_KEY>
Result returnSynchronous; the result is contained in the response body
Quality tier qualityauto / low / medium / high / xhigh / max
Planning tier complexitylow / medium / high, default high
Output formatspng / jpeg / webp / gif / svg
Guardrail removal supportedNot supported

What Changed Versus 2.5 โ€‹

ChangeDescription
Generation speedAbout 50% faster overall. Flare reaches roughly 6 seconds per image at the low tier
complexity decoupled from qualityA new complexity parameter independently controls "planning effort before rendering", separate from quality, which governs pixel detail
Animated GIF outputoutput_format=gif, supporting 2~16 frames with adjustable per-frame duration
SVG vector outputoutput_format=svg, producing infinitely scalable vector graphics directly
Markup-based editingCarries over the annotation-layer capability from 2.5

Choosing Between complexity and quality

  • complexity governs how much compute the model spends planning the scene before it starts drawing; quality governs pixel rendering precision.
  • Local tweaks (swapping a background, changing a signboard) โ†’ complexity=low plus whatever quality you need, which skips re-laying-out the scene.
  • Many elements in one frame, dense text, complex layouts โ†’ raise complexity.
  • Changing complexity is not billed separately, but higher tiers add latency and may introduce unintended changes. Benchmark it on your own workloads.

Submit a Text-to-Image Task โ€‹

1. Endpoint Description โ€‹

  • Request domain (international site): tokenhub-intl.tencentcloudmaas.com (Singapore)
  • Endpoint: POST /v1/wand/og-image/generation

Generates an image synchronously from the input text and returns the result directly.

2. Input Parameters โ€‹

ParameterRequiredTypeDescription
modelYesStringModel selection: custom-model-og-v2.6-f or custom-model-og-v2.6-s
promptNoStringText description. Up to 32,000 characters
sizeNoStringOutput image size. Custom sizes are supported (each side โ‰ค 3840px, width and height multiples of 16, long-to-short side ratio โ‰ค 3:1, total pixels 655,360~8,294,400)
qualityNoStringQuality tier. auto, low, medium, high, xhigh, max
complexityNoStringPlanning effort before rendering, independent of quality. low / medium / high, defaults to high when omitted. Not billed separately
nNoIntegerNumber of images to generate, 1~10. Default: 1. Must be 1 for SVG output
backgroundNoStringBackground transparency. opaque, transparent, auto. When transparent is chosen, the output format must be png or webp
output_formatNoStringOutput format. png (default), jpeg, webp, gif, svg
output_compressionNoIntegerCompression level (0-100). Supported only for webp / jpeg output, default 100
frame_countNoIntegerGIF frame count. 2~16, defaults to 9 when omitted. Supported only when output_format=gif
frame_duration_msNoIntegerGIF per-frame duration (milliseconds). Minimum 20, defaults to 200 when omitted. Rounded to the nearest 10ms during encoding; the per-frame cap is 655,350ms. Supported only when output_format=gif
svgNoObjectSVG author-model settings. Requires output_format=svg. See "SVG Vector Output" below
response_formatNoStringReturn format of the generated image. Options: url or b64_json

3. Output Parameters โ€‹

ParameterTypeDescription
output_formatStringReturn format of the generated image. Example: png
qualityStringQuality actually applied. Example: medium
sizeStringSize of the generated image. Example: 1024x1024
backgroundStringBackground setting actually applied. Example: auto
createdIntegerUnix timestamp (in seconds) of image creation. Example: 1776946569
dataArray of ResultImageList of generated result images
usageUsageToken statistics
request_idStringUnique request ID generated by the server and returned with every request. Provide this RequestId when troubleshooting

4. Examples โ€‹

Request Example

bash
curl --location 'https://tokenhub-intl.tencentcloudmaas.com/v1/wand/og-image/generation' \
--header 'Authorization: Bearer YOUR_API_KEY' \
--header 'Content-Type: application/json' \
  -d '{
    "model": "custom-model-og-v2.6-f",
    "quality": "medium",
    "complexity": "low",
    "prompt": "a dancing kitten",
    "size": "1024x1024",
    "n": 1
}'

Response Example

json
{
    "background": "opaque",
    "created": 1786095835,
    "data": [
        {
            "url": "https://xxxxxxxxxxxx.png"
        }
    ],
    "output_format": "png",
    "quality": "medium",
    "size": "1024x1024",
    "usage": {
        "total_tokens": 27180
    },
    "request_id": "eaa4b07b-ccde-40a4-b184-822e3c5a4e60"
}

Animated GIF Output โ€‹

Set output_format to gif to produce an animation directly โ€” no separate video model and no manual frame stitching required.

Parameters โ€‹

ParameterValuesDefaultDescription
output_formatgifโ€”Required; enables GIF output
frame_count2~169Number of frames
frame_duration_msโ‰ฅ 20200Per-frame duration (milliseconds), rounded to the nearest 10ms

At the defaults, 9 frames ร— 200ms โ‰ˆ a 1.8-second loop.

Limitations โ€‹

LimitationDescription
Supported endpointsText-to-image endpoint only; not supported on image editing endpoints
Unsupported capabilitiesStreaming responses, partial_images, mask, custom output_compression
Size constraintsThe requested size must satisfy the selected model's GIF size limits
EntitlementRequires the model's GIF preview entitlement

Example โ€‹

bash
curl --location 'https://tokenhub-intl.tencentcloudmaas.com/v1/wand/og-image/generation' \
--header 'Authorization: Bearer YOUR_API_KEY' \
--header 'Content-Type: application/json' \
  -d '{
    "model": "custom-model-og-v2.6-f",
    "prompt": "an owl blinking, seamless loop",
    "output_format": "gif",
    "frame_count": 12,
    "frame_duration_ms": 120,
    "size": "1024x1024"
}'

SVG Vector Output โ€‹

Set output_format to svg to obtain infinitely scalable vector graphics directly โ€” ideal for icons, illustrations, and diagrams, and it removes the raster-to-vector conversion step.

How It Works โ€‹

The server first generates a reference image with the image model, then an SVG author model produces the vector content from it. Consequently:

  • The top-level model determines which image model generates the reference image;
  • In the SVG case, complexity controls the planning effort for the reference image, independent of the SVG author model's reasoning effort;
  • svg.model specifies the SVG author model, and svg.reasoning controls its reasoning effort.

Parameters โ€‹

ParameterTypeDescription
output_formatStringSet to svg
svg.modelStringThe SVG author model. Required when author-model selection is enabled; otherwise the entire svg object should be omitted
svg.reasoningObjectReasoning settings for the SVG author, independent of the image complexity. When omitted or null, the author model's default is used

Limitations โ€‹

LimitationDescription
Supported endpointsText-to-image endpoint only; not supported on image editing endpoints, and SVG cannot be used as an input image
nMust be 1 โ€” one file per request
Unsupported capabilitiesStreaming responses, partial_images, prompt_rewriter, output_compression
OutputStatic vector graphics

Example โ€‹

bash
curl --location 'https://tokenhub-intl.tencentcloudmaas.com/v1/wand/og-image/generation' \
--header 'Authorization: Bearer YOUR_API_KEY' \
--header 'Content-Type: application/json' \
  -d '{
    "model": "custom-model-og-v2.6-s",
    "prompt": "a minimalist weather icon set with linear strokes",
    "output_format": "svg",
    "n": 1
}'

SVG Is Billed Differently

SVG output incurs two charges at once: the image model bills the reference image's tokens at its own rate, and the SVG author model bills the tokens used to produce the vector content at its rate. The total cost is the sum of the two.

Submit an Image Editing Task โ€‹

1. Endpoint Description โ€‹

  • Request domain: tokenhub-intl.tencentcloudmaas.com (Singapore)
  • Endpoint: POST /v1/wand/og-image/edits (json)
  • Or: POST /v1/wand/og-image/multipart-edits (multipart)
  • Or: POST /v1/wand/og-image/images/edits (json / multipart)

Generates an image synchronously from the input image and returns the result directly.

2. Input Parameters โ€‹

ParameterRequiredTypeDescription
modelNoStringModel selection: custom-model-og-v2.6-f or custom-model-og-v2.6-s
promptYesStringText description. Up to 32,000 characters
imagesYesArray of ImageRefInput images. Supported count: 1~16
maskNoImageRefUsed for partial regeneration (inpainting); must match the input image size
markupNoImageRefAnnotation layer. A transparent PNG that circles the region to change. Cannot be used together with mask
sizeNoStringOutput image size. Custom sizes are supported (each side โ‰ค 3840px, width and height multiples of 16, long-to-short side ratio โ‰ค 3:1, total pixels 655,360~8,294,400)
nNoIntegerNumber of images to generate, 1~10. Default: 1
qualityNoStringQuality tier. auto, low, medium, high, xhigh, max
complexityNoStringPlanning effort. low / medium / high, default high. Use low for local edits
backgroundNoStringBackground transparency. transparent, opaque, or auto (default). Example: auto
output_formatNoStringOutput format. Supports png, jpeg, webp. The editing endpoints do not support gif or svg
output_compressionNoIntegerCompression level (0-100). Supported only for webp / jpeg output
response_formatNoStringReturn format of the generated image. Options: url or b64_json

transparent Is Available Only in PNG / WebP

For a transparent background, use background=transparent together with output_format=png or webp.

3. Markup-Based Editing โ€‹

Use an annotation layer to tell the model where to edit โ€” far more precise than describing the location in words ("move the logo a little to the left").

ConstraintDescription
TargetThe markup applies to the first input image only; the remaining inputs act as reference images
File formatMust be a standalone transparent PNG containing transparent pixels; do not flatten the annotation into the source image
Size requirementThe markup layer must match the first input image in width and height exactly
Pixel capThe first input image and the markup layer must each stay under 16,777,216 pixels
File sizeThe markup file must not exceed 50 MB
Mutually exclusiveCannot be used together with mask
Supported endpointsImage editing endpoints only; not supported on the text-to-image endpoint

Request Example (multipart)

bash
curl --location 'https://tokenhub-intl.tencentcloudmaas.com/v1/wand/og-image/multipart-edits' \
--header 'Authorization: Bearer YOUR_API_KEY'  \
--form 'model=custom-model-og-v2.6-f' \
--form 'quality=medium' \
--form 'complexity=low' \
--form 'image[]=@kitchen.png' \
--form 'markup=@markup-overlay.png' \
--form 'prompt=change the striped vase inside the red circle to solid cobalt blue ceramic'

4. Output Parameters โ€‹

Identical to the text-to-image endpoint; see above.

5. Examples โ€‹

Request Example (json)

bash
curl --location 'https://tokenhub-intl.tencentcloudmaas.com/v1/wand/og-image/edits' \
--header 'Authorization: Bearer YOUR_API_KEY'  \
--header 'Content-Type: application/json' \
  -d '{
    "model": "custom-model-og-v2.6-s",
    "quality": "high",
    "complexity": "low",
    "prompt": "dancing",
    "images": [{ "image_url": "https://xxxxxxxxxxxx.png"}],
    "size": "1024x1024",
    "n": 2
}'

Response Example

json
{
    "background": "opaque",
    "created": 1786096322,
    "data": [
        {
            "url": "https://xxxxxxxxxxxx.png"
        },
        {
            "url": "https://xxxxxxxxxxxx.png"
        }
    ],
    "output_format": "png",
    "quality": "high",
    "size": "1024x1024",
    "usage": {
        "total_tokens": 65920
    },
    "request_id": "705d12d2-9f68-4cfa-859a-733e449ea675"
}

Capability Support Matrix โ€‹

CapabilityText-to-image endpointImage editing endpoints
complexityโœ…โœ…
markup annotation editingโŒโœ…
mask partial regenerationโŒโœ…
Animated GIF outputโœ…โŒ
SVG vector outputโœ…โŒ

Data Structures โ€‹

ResultImage โ€‹

ParameterTypeDescription
urlStringURL of the returned image

Usage โ€‹

ParameterTypeDescription
total_tokensIntegerToken statistics

ImageRef โ€‹

NameTypeRequiredDescription
image_urlStringNoImages support png and jpg. Pass either the image URL or base64-encoded data. Example: https://xxxxxxxxxxx.png

Billing and Pricing โ€‹

  • Billing mode: Pay-as-you-go (by token), postpaid.
  • Settlement cycle: The system aggregates and settles your resource consumption hourly.
  • Pricing logic: Cost is determined by token usage and the corresponding unit price: total cost = token usage ร— token unit price.
  • Usage notes: A unified unit price of 1.6 USD per million tokens applies. The conversion against upstream vendor tokens is shown below.
Token typeVendor USD unit priceTokenHub priceConversion factor
text input$5.00 / million$1.60 / million3.125
text cached$1.25 / million$1.60 / million0.78125
image input$8.00 / million$1.60 / million5.0
image cached$2.00 / million$1.60 / million1.25
image output$30.00 / million$1.60 / million18.75

USD Settlement Logic (International Site)

text input tokens ร— 3.125
+ text cached input tokens ร— 0.78125
+ image input tokens ร— 5
+ image cached input tokens ร— 1.25
+ output image tokens ร— 18.75
= token consumption of a single task

The Real Cost of 2.6 Differs From 2.5

The token unit price is identical to 2.5, but the output image token count rises significantly at the low and medium tiers, so the per-image cost at those two tiers goes up. Token consumption at high and above is unchanged from 2.5.

Taking 1024x1024 as an example, output image token counts compare as follows:

quality2.52.6Change
low196823about 4.2ร—
medium4391,449about 3.3ร—
high1,7561,756Unchanged
xhigh3,1223,122Unchanged
max7,0247,024Unchanged

The token counts above are reference values, derived from OpenAI's official token calculator at 1024x1024. Actual consumption is whatever the endpoint returns in usage.total_tokens.

Selection advice: If your workload is dominated by low-quality batch generation (icons, thumbnails, bulk drafts), per-image cost on 2.6 will be noticeably higher than on 2.5 and your budget needs recalculating. At high and above, 2.6 is "same price, faster" and can be adopted directly.

DocumentDescription
GPT-image 2.0The base version, with quality tiers up to high
GPT-image 2.5Adds the xhigh / max quality tiers and markup-based editing
Get an API KeyApplying for and managing TokenHub API Keys

Built with VitePress ยท Deployed on Tencent EdgeOne Pages