GPT Image 2.6

1 models available

Provider: OPENAI

Updated: Recently

GPT Image 2.6 is a new-generation image generation model released by OpenAI on November 4, 2026. It consists of two sub-models: Flare (Speed Version) and Sunburst (High-Precision Version). Based on GPT Image 2.5, it adds GIF and SVG image output formats, with its inference speed doubled!

Available Models

GPT Image 2.6 Flare

gpt-image-2.6-flare

GPT Image 2.6 Flare is OpenAI's image generation and editing model optimized for speed: fast iteration and interactive text-to-image and image-to-image workflows, generating about 50% faster than GPT Image 2.5. It takes text and image inputs and produces images, with quality tiers from low to max, transparent backgrounds, and custom sizes in multiples of 16 pixels at aspect ratios from 1:3 to 3:1. GPT Image 2.6 introduces a complexity setting for how much the model plans before rendering, markup-guided edits, animated GIF output, and SVG output. Choose Sunburst instead when detail and edit-to-edit consistency matter more than latency.

OpenAI · GPT Image 2.6

Released November 4, 2026 · Two tiers: Flare and Sunburst

One prompt. An actual GIF and SVG.

For the first time, an image model delivers motion and vectors — not just a flat bitmap.

GPT Image 2.6 keeps everything from 2.5 and adds looping GIF animation, editable SVG vector output, a complexity control and markup-guided editing. Flare trades compute for iteration speed; Sunburst trades latency for production detail. Both share one request shape and their own public model IDs.

Get API access Read the API docs OpenAI guide

Output

5 formats

PNG / JPEG / WebP / GIF / SVG, switched with a single output_format field.

GIF frames

2–64

Paired with a 20–2000ms per-frame duration to control loop rhythm.

Model tiers

2 tiers

Flare optimizes for speed, Sunburst for precision; identical parameters.

Planning depth

3 levels

low / medium / high control how much the model plans before rendering, at no extra charge.

Example prompt

A monkey in a green coat lifts a glowing lantern in front of a snowy cabin, blinks, then lowers it again. The motion returns to the starting pose so the clip loops seamlessly; the warm lantern light spills softly across the snow, and the background stays quiet.

Official example: Sunburst generated this lantern animation with frame_count=9 and frame_duration_ms=200 — 1.8 seconds that loop seamlessly.

One field, four deliverables

output_format decides what comes back. Bitmaps suit photography and compositing, GIF suits social and micro-interactions, SVG suits logos, icons and illustrations you need to edit again.

Lossless bitmap

PNG

Keeps the alpha channel — the default for transparent product cutouts and the safest path for editing and compositing.

Fastest bitmap

JPEG

The quickest and smallest bitmap option for latency-sensitive batches. No transparency.

Web bitmap

WebP

Balances size and quality, supports transparency and a custom output_compression value.

Looping animation

GIF

A complete looping animation from one prompt, returned as b64_json you can write straight to a .gif file.

Editable vector

SVG

Static vector XML with editable paths, shapes and colors. No scripts or external resources.

GIF and SVG are only available from the generation endpoint — image edits return neither — and both ignore output_compression.

GIF animation: describe the motion

Set output_format to gif, then describe the subject, the motion and how the loop should close. What comes back is a finished animation, not a still you have to frame yourself.

Official example prompt (translated)

Three fields drive the animation

frame_count2–64 frames: Total frames. More frames mean smoother motion, but frames × frame size × n must fit the model's pixel budget.
frame_duration_ms20–2000ms: How long each frame holds, rounded to 10ms. Nine frames at 200ms is roughly a 1.8-second loop.
sizeauto / multiples of 16: GIF frame size. Optional — leave it empty to derive it from resolution + aspect_ratio. Because GIF frames are budgeted together (frames x pixels must fit the per-request limit), APIPod automatically scales the frame size down as frame_count grows, keeping the aspect ratio and 16-alignment.

Writing prompts that loop

Subject: pin one moving subject — a character, object or scene. Competing motions muddy the animation.
Motion: describe two to four key poses in order, e.g. "lifts → blinks → lowers".
Loop: ask explicitly for the end pose to match the start, so the clip repeats without a jump.
Light: motion blurs light more easily than stills. Naming the light direction and color temperature lifts the hit rate.

Limits

· GIF is generation-only: no reference-image edits, no masks.
· No streaming, partial images, custom output_compression or prompt_rewriter.
· Transparency applies to bitmaps only — GIF has no alpha channel.

SVG vectors: shippable and editable

Set output_format to svg and the model renders a reference bitmap, then converts it to vectors. The result is plain static XML — no scripts, external resources or embedded rasters — ready for a design tool or your front end.

GPT Image 2.6 Flare — Official example prompt (translated)
Create an original, non-infringing logo for Field & Flour, a local bakery. It should feel warm, simple and timeless: clean vector-like shapes, a strong silhouette, balanced negative space. Favor simplicity over detail so it reads at both small and large sizes. Flat design, minimal strokes, no gradients unless essential. Fully transparent background, a single centered logo with generous padding, clean alpha edges, and no solid backdrop, scenery, checkerboard or watermark.

GPT Image 2.6 Sunburst — Official example prompt (translated)
Create an original, non-infringing logo for Field & Flour, a local bakery. It should feel warm, simple and timeless: clean vector-like shapes, a strong silhouette, balanced negative space. Favor simplicity over detail so it reads at both small and large sizes. Flat design, minimal strokes, no gradients unless essential. Fully transparent background, a single centered logo with generous padding, clean alpha edges, and no solid backdrop, scenery, checkerboard or watermark.

When to choose SVG over PNG

Logos, icons and brand marks: recoloring, scaling and deconstructing all happen in vectors.
Inline front-end use: an SVG can ship as a component and weighs far less than a bitmap.
Print and large-format work: vectors stay crisp at any size.

Parameter rules

n must be 1 — one SVG per request.
output_compression is not supported; compression is meaningless for vectors.
The edit endpoint does not produce SVG, and an SVG is not a valid raster input for edits.
quality controls the reference bitmap, which directly shapes how much detail survives the vector conversion.

Markup-guided edits: show it, don't just describe it

Overlay an annotation layer on a reference image — a circle, box or arrow — then describe the change in the prompt. The model locks onto the region more reliably and leaves the rest of the frame untouched.

Official example prompt (translated)

Recolor the striped vase inside the red circle to solid cobalt blue ceramic. Keep its shape, the dried stems and the rest of the kitchen unchanged. Remove the red circle from the finished photo.

1
Prepare a clean reference image and a transparent PNG overlay that circles the target region.
2
Submit both as multipart files: image plus markup, which travels on its own channel.
3
In the prompt, state what to change and what must stay, and ask for the annotation to be removed.

complexity: how long the model thinks first

complexity controls how much the image model plans before rendering. It is a separate axis from quality: quality decides how finely it renders, complexity decides how deeply it reasons. Changing it costs nothing extra.

low
Focused edits with one clear target — swap a background, repaint a sign. Faster and cheaper.

medium
The balance point for everyday campaign assets and multi-element compositions.

The same prompt can compose differently at different complexity levels — in OpenAI's example, low kept only the trophy and banner while high filled in the jerseys and the stands. Compare results when switching; higher is not automatically better.

One family, two tiers

Flare and Sunburst take exactly the same parameters. The difference is how much time the model is willing to spend.

FlareRecommended

Speed tier

The default when response time matters: fast iteration, batch variants and in-product generation.

Lower per-request latency for tight loops
Supports all five output formats
Public model ID: gpt-image-2.6-flare

View Flare

SunburstPrecision flagship

Detail tier

The choice when detail is worth waiting for: fine visual work, multi-round edits and production-ready assets.

Stronger detail retention and edit consistency
Supports all five output formats
Public model ID: gpt-image-2.6-sunburst

View Sunburst

Request fields at a glance

GPT Image 2.6 keeps the GPT Image 2 request shape and adds three fields. Migrating is a model-ID change.

Field	Values	Notes
model	gpt-image-2.6-flare / gpt-image-2.6-sunburst	The public model ID that picks the tier.
prompt	≤ 4000 characters	Scene description; for GIF, include subject, motion and the loop target.
output_format	png / jpeg / webp / gif / svg	New in 2.6: gif and svg. Both are generation-only.
quality	low / medium / high / xhigh / max	Fidelity tier; omitted or auto is sent upstream as high.
complexity	low / medium / high	New in 2.6: planning depth; defaults to high.
frame_count / frame_duration_ms	2–64 / 20–2000ms	New in 2.6, only read when output_format=gif.
resolution + aspect_ratio	1K / 2K / 4K × 15 ratios	Converted to upstream pixel sizes.
size	auto, or WIDTHxHEIGHT (multiples of 16)	GIF frame size. Optional — leave it empty to derive it from resolution + aspect_ratio. Because GIF frames are budgeted together (frames x pixels must fit the per-request limit), APIPod automatically scales the frame size down as frame_count grows, keeping the aspect ratio and 16-alignment.
moderation	low / auto	Moderation strictness. APIPod defaults to low to reduce false blocks; pass auto for OpenAI's default filtering.
background	auto / opaque / transparent	Transparent needs quality ≥ medium and only works through reference-image edits.

A few lines to integrate

GIF, SVG and bitmaps share the same /v1/images/generations endpoint — swap output_format.

generate.py

import base64, requests

resp = requests.post(
    "https://api.apipod.ai/v1/images/generations",
    headers={"Authorization": "Bearer $API_KEY"},
    json={
        "model": "gpt-image-2.6-sunburst",
        "prompt": "A monkey in a green coat lifts a glowing lantern, blinks, then lowers it; the loop returns to the starting pose.",
        "output_format": "gif",
        "frame_count": 9,
        "frame_duration_ms": 200,
        "quality": "high",
    },
)
data = resp.json()["data"][0]["b64_json"]
open("lantern.gif", "wb").write(base64.b64decode(data))

The response's data[0].b64_json is the base64 payload: write it to .gif, .svg or the bitmap's own extension.

Get an API key

GIF animation

"output_format": "gif"

SVG vector

"output_format": "svg"

Bitmap

"output_format": "png"

FAQ

How GPT Image 2.6 works on APIPod.

What is GPT Image 2.6?+

GPT Image 2.6 is the next generation of OpenAI's image family, released November 4, 2026 as two tiers, Flare and Sunburst. On top of 2.5 it adds native GIF animation output, native SVG vector output and a complexity control; the edit path also accepts markup overlays.

How do I generate a GIF?+

Set output_format to gif on /v1/images/generations and describe the subject, motion and loop target in the prompt. Optionally set frame_count (2–64, default 9) and frame_duration_ms (20–2000, default 200) to control the rhythm. Decoding data[0].b64_json gives you the finished GIF. GIF does not support image edits, streaming or custom compression.

Is SVG output reliable enough to use?+

The model renders a reference bitmap and converts it to vectors, so the quality tier directly shapes how much detail survives. The result is static XML with no scripts, external resources or embedded rasters — good for logos, icons and illustrations. The edit endpoint does not return SVG.

Why did my GIF come back as a PNG?+

Check three things: output_format is explicitly set to gif; the request did not go through the image edit endpoint (GIF is generation-only); and data[0] returns b64_json rather than a URL. APIPod sniffs the artifact header and stores it with the correct Content-Type, so a PNG result normally means the parameter never reached the upstream.

How is complexity different from quality?+

quality controls rendering fidelity (low / medium / high / xhigh / max); complexity controls how much the model plans before rendering (low / medium / high). The first affects fine detail, the second affects composition completeness and latency. Adjusting complexity costs nothing extra, and it defaults to high.

What is markup-guided editing?+

You overlay a transparent annotation layer on a reference image — circles, boxes or arrows marking the target — and describe the change in the prompt. The edit endpoint receives the overlay on its own channel, which lets the model lock onto the region while leaving the rest untouched and removing the annotation marks.

How are GIF and SVG billed?+

With the same token model as bitmaps: a hold at submission and a settle-up on completion, refunding the difference. GIF frame count and dimensions, and the SVG's reference image plus vector tokens, all show up in the final usage — more frames or larger sizes cost more. Check the pricing page for current rates.

What changes when migrating from GPT Image 2.5?+

Only the model ID: gpt-image-2.5-flare becomes gpt-image-2.6-flare, and the same for sunburst. Existing parameters, endpoints and response fields are compatible; add output_format=gif/svg, complexity or the frame_* fields when you want the new capabilities.

Start shipping with GPT Image 2.6

Bitmaps, animation and vectors — one API key, one request shape.

Start free Talk to sales