GPT Image 2.6 Released: Speed, Price, and SVG Output Changes

GPT Image 2.6 is now officially released. OpenAI's API changelog dates the release as November 4, 2026, while the official developer account announcement was posted on the morning of November 5 Beijing time. API changelog and official announcement contain the specifics for both models.

Two models: one for detail, one for speed

The new models continue to be split into Sunburst and Flare. Both accept text and image input for generation and editing. Sunburst targets compositions and edits with more detail; Flare targets interactive scenarios needing rapid iteration. The actual API IDs are gpt-image-2.6-sunburst and gpt-image-2.6-flare — don't confuse product names with model IDs when integrating.

For multi-turn modifications to the same hero visual, evaluate Sunburst first; when you need to quickly try different concepts, evaluate Flare first. Both have the same official per-token price, so selection should focus on result quality and response speed for the specific task. Sunburst model page, Flare model page

Compared to 2.5: what changes have been confirmed

Aspect	GPT Image 2.5	GPT Image 2.6
Model branches	Sunburst, Flare	Continues Sunburst, Flare
Generation speed	Previous-gen baseline for this announcement	Official claim: 50% faster; Flare at low quality ~6 seconds
Continuous editing	Existing image editing capability	Official emphasis: Sunburst maintains better detail consistency at medium and high quality
Output capabilities	Existing image generation and editing	Adds SVG and animated GIF generation
Pre-generation planning	Shouldn't directly apply new params	Adds complexity control
Markup-guided editing	Supported on applicable 2.5 models	Supported; newly opened interface capability
Standard token price	Same as right column	No increase with this upgrade

Speed figures come from the official X announcement, not a promise of latency for any prompt, resolution, or service load. Actual time varies with task complexity and runtime conditions. The continuous-editing improvement is also conditional — the announcement specifically emphasizes Sunburst at medium and high quality.

SVG, GIF, complexity, and markup support scope can be verified in the official changelog. Markup-guided editing applies to both supported 2.5 and 2.6 models; it isn't exclusive to 2.6.

Pricing unchanged from 2.5, per token

The official standard API prices below are USD per million tokens. Sunburst and Flare use the same rates within the same generation.

Billing item	2.5 Sunburst / Flare	2.6 Sunburst / Flare
Text input	$5	$5
Cached text input	$1.25	$1.25
Image input	$8	$8
Cached image input	$2	$2
Image output	$30	$30

2.6 continues 2.5's standard token rates; actual cost per image depends on token usage. The number of text and image input tokens and the output tokens consumed all affect the bill; resolution and quality differ, so you can't just compare cost by image count. Official pricing table

If a batch of requests totals 1 million image output tokens, that output alone costs $30, with input billed separately. When evaluating migration, record success rate, completion time, and total cost for the same task set, then compare what one usable image actually costs.

Focus: SVG output

For web and design asset production, SVG output enables more convenient downstream editing. It lets generated results retain paths, shapes, and colors for continued editing. Icons, simple logos, and illustrations are the official use cases.

Previously, getting a raster image and wanting to change a local color or shape often meant regenerating or vectorizing separately. Now the API offers a direct SVG output endpoint. Generated assets can be further adjusted for color, shape, and path. Whether the model's output paths are clean and groupings are convenient still needs manual inspection; the official line doesn't promise it equals a hand-curated design source file.

The official description says the process first generates a reference image then produces SVG from it; quality controls the reference image quality. The returned content is static vector, with no scripts, external resources, or embedded bitmaps. It also has clear interface boundaries: SVG only works on the generation endpoint, n must be 1, no streaming, and the editing endpoint cannot output SVG. SVG output notes

The example below shows the basic structure of an SVG generation request. At runtime, the API key is read from an environment variable; costs are based on actual calls.

curl https://api.openai.com/v1/images/generations \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.6-sunburst",
    "prompt": "Design a simple vector icon for a bakery using bread and wheat elements, clean outlines, limited to three colors, no text",
    "output_format": "svg",
    "quality": "medium",
    "n": 1
  }' > response.json

The returned data[0].b64_json must be decoded and saved as .svg, not merely renamed. After saving, inspect paths, colors, scaling behavior, and file size in a browser or vector editor before using it in production. The API reference also lists svg_usage separately to record model usage for SVG generation, distinct from image rendering usage. Therefore, the image token rate above can't be taken as a guarantee of total SVG request cost; check actual usage and the corresponding model rate when budgeting. Interface parameters follow the image generation API reference.

For blog illustrations, start with simple icons or illustrations that have few structural elements and little text. For technical diagrams with many Chinese labels, strict alignment, and precise dimensions, still proofread item by item. SVG provides editing room, but content correctness still requires human review.

GIF and editing control also worth knowing

The official announcement demonstrated turning a static character into a looping GIF, e.g., waving then returning to the starting pose. ChatGPT release notes confirm GIF creation is rolling out gradually, with adjustable playback speed and shareable results. Official example announcement, ChatGPT release notes

The API also offers complexity, which controls the degree of pre-generation planning, separate from the rendering quality. For local edits, OpenAI recommends starting with low complexity, then trying medium or high for more complex scenes; not everything needs to be maxed out. Official editing advice

Cases on X: what progress is visible from the images

Following the release posts on X, I found three comparison sets from DesignArena and an SVG demo from OpenAI Developers. The commentary below covers the public examples image by image; I did not call the API to run benchmarks or re-time generation. The 'Winner' labels are the publisher's annotations, not a blind test reorganized in this article.

Birthday cake: fewer extra slogans is more practical than extra detail

DesignArena's cake comparison used a short prompt: generate a beautiful birthday cake. The left is 2.6 Flare, labeled High; the right is 2.5 Sunburst, with no quality setting listed for it.

Both images already have respectable cream, strawberries, and candlelight. The difference is in the background: 2.5 added two cards with blessing text, 2.6 didn't. For product imagery, this tendency to omit irrelevant elements is useful, saving you from erasing unwanted text later.

However, 2.6's cake still has a birthday greeting written on it. This image supports "fewer extra slogans", not "no text is generated when no text is requested". And this is a single cross-branch pair, so it can't support a verdict that 2.6 is comprehensively better than 2.5. For a retest, fix quality, size, and prompt, then compare within the same branch.

Rabbit picture book: switching to dusk, character and scene still recognizable

The second set is an image edit: input a daytime rabbit illustration, ask to change it to dusk, adding warm flashlight glow, night fog, and stars. The comparison model is FLUX.2 Klein 4B Distilled Edit, not GPT Image 2.5.

2.6's result preserves the blue floral dress, light apron, the posture looking down at the flashlight, and the general position of the rabbit house on the right. The right-side result turns the rabbit to face forward, with notable changes in clothing and body pose. For tasks like picture books and character posters that need to reuse a character, the former is closer to continuing to edit the original image.

But "keeping the character" has degrees. The input thumbnail in the public image is small and doesn't allow pixel-level confirmation of dress patterns, flowers, or architectural details. My judgment is that general identity and composition are preserved better; to verify whether multi-turn edits stay consistent, you'd need to modify several times, and this set only shows one turn.

Old warehouse: was the complex requirement drawn in?

The third warehouse scene compares 2.6 Sunburst with Microsoft's MAI-Image-2.6 Flash. Both have "2.6" in the name but belong to different vendors. The image shows GPT settings quality: auto, complexity: high.

The left image shows dust and mist visible through the top light, a foreground old painting leaning against a damaged wall, and both canvas and wall showing damp, peeling texture. The right image also achieves a decayed atmosphere, but the boxes and walls carry more text. Rather than comparing which is more cinematic, I care whether the objects, materials, and air effects specified in the prompt were actually rendered, and whether the model added extras.

This image shows Sunburst performs well on this particular complex scene, but it doesn't prove high complexity is always better: the publisher didn't provide a same-model comparison across low, medium, and high complexity, nor did they provide usage and latency for the request.

Geometric deer SVG: shows the use case, doesn't prove the file is easy to edit

OpenAI Developers' release video, around the 55-second mark, shows a deer made of black geometric blocks to illustrate SVG support.

Clean outlines, few colors, and no complex textures make this kind of graphic suitable for observing the vector output use case. For actual delivery, you'd open the source file: are the antlers and body easy to select separately? How many paths to change a color? When scaled down to an icon, do thin gaps disappear?

The video image can't answer those questions, so this counts as a feature demo, not an evaluation of path quality or editing efficiency. The screenshot itself is a raster, not the SVG source shown in the video.

How I'd choose a model

From these examples, I'd try 2.6 first for two kinds of work: reducing extra elements in product images, and adjusting lighting or scene on an existing character. For the former, start with Flare; for prompts with many requirements and detail, try Sunburst. For SVG, start with simple logos or icons, then inspect the source file before using it.

Speed and cost need separate measurement. The comparison images above have no per-request timing or billing, so they can't validate "50% faster" or "cheaper per image". To decide whether to migrate from 2.5, use your own fixed batch of assets to record three things: time to first usable result, number of revision rounds, and total cost.

Where to use it now

Entry point	Confirmed info	Notes
ChatGPT	Images 2.6 announced; GIF feature rolling out gradually	Release notes don't give uniform quotas for all plans; features and limits depend on your account
Codex	Official developer announcement confirms 2.6 and introduces in-image overlay editing	Exact entry points depend on client and account display
OpenAI API	Both models now listed on model pages and changelog	Use correct model IDs; new capabilities follow the Image API docs; GIF API reference also requires the corresponding preview permission

ChatGPT also adds a save-reference-image feature, letting you reference the image by name in later chats and reducing re-uploads. Saved reference images are a ChatGPT product feature; API image input and context management are handled separately via interfaces. ChatGPT release notes

If you already have 2.5 applications, compare both versions with the same set of inputs before migrating. Especially keep intermediate results from continuous edits and check whether character clothing, product labels, and background details drift across modifications. The officially claimed consistency gains ultimately need to be verified against your own assets.