OpenAI released GPT-Image-2.6 on November 4 and did something the image API has never done before: split generation into two purpose-built models. Flare is the speed model. Sunburst is the quality model. Both cost the same. That is a meaningful architectural choice — but it is not the only change worth your attention. The release also adds SVG export, animated GIF output, and a new complexity parameter. It quietly drops Batch API support. If you are running GPT-Image-2.5 in production, that last item deserves a closer look before you upgrade.
Flare vs Sunburst: Make the Call Once, Not Every Request
Both models share the same endpoints, the same pricing ($5 per million text input tokens, $8 for image input, $30 for image output), and the same controls. The only difference is what they optimize for.
Flare is built for iteration. At low quality, it can return an image in roughly six seconds — fast enough for a generation loop or a live-preview workflow. It suits rapid variation cycles: generating twenty layout options, testing copy treatments, prototyping ad creatives.
Sunburst is built for consistency. It holds up better across repeated edits and produces sharper detail at medium and high quality settings. Use it when the image is heading to production: final marketing assets, product photography composites, anything a human is about to approve.
The practical workflow: prototype with Flare, finalize with Sunburst. You are not locked in per-request — just pick the right model for the stage of work. OpenRouter’s side-by-side comparison shows the output quality differences in practice.
SVG Output: Vector Generation Arrives, With a Billing Catch
GPT-Image-2.6 can now output native SVG. This is genuinely useful for icon generation, logo drafts, and scalable UI assets that need to work at any resolution.
The catch: SVG generation runs two models per request. The image model renders a reference raster image, then a separate authoring model (currently gpt-6.1-sol) writes the vector code. Both are billed. Budget accordingly — SVG requests cost more than standard PNG generations.
response = client.images.generate(
model="gpt-image-2.6-flare",
prompt="A minimal icon for a terminal application",
output_format="svg",
n=1,
extra_body={"svg": {"model": "gpt-6.1-sol"}}
)
# Decode: base64.b64decode(response.data[0].b64_json)A few hard constraints: n must be 1, streaming is not supported, and prompt_rewriter is disabled. The output is base64-encoded UTF-8 XML — decode and save as .svg. For a deeper look at the dual-billing mechanics, this breakdown of GPT-Image-2.6’s hidden limits covers each constraint in detail.
Animated GIF Output: One Call Replaces the Stitch Workflow
Previously, generating animated imagery meant creating frames individually and stitching them in post. GPT-Image-2.6 collapses that into a single API call.
response = client.images.generate(
model="gpt-image-2.6-sunburst",
prompt="A spinning globe with orbiting tech icons",
output_format="gif",
extra_body={
"gif": {
"frame_count": 12,
"frame_duration_ms": 150
}
}
)frame_count accepts 2 to 64 frames (default: 9). frame_duration_ms accepts 20 to 2000ms in 10ms increments (default: 200ms). The defaults produce a 1.8-second loop. At 64 frames with 2000ms intervals, you get over two minutes of animation — though at that point you are paying for a substantial number of image output tokens.
One constraint worth knowing: GIF output only works on the /v1/images/generations endpoint. The edits endpoint does not support it. You cannot animate an existing image; you can only generate animated content from scratch.
The Complexity Parameter: Not the Same as Quality
GPT-Image-2.6 adds a complexity parameter alongside the existing quality parameter. These are not interchangeable. Quality controls rendering fidelity. Complexity controls how much the model plans the composition before it starts rendering.
- low: fast and targeted — best for single-element edits (swapping a background, changing sign text, recoloring an object)
- medium: balanced for moderate compositional work
- high (default): needed for dense text layouts, scenes with many interacting objects, or complex spatial relationships — but it adds latency
OpenAI’s own guidance is to default to low complexity for most editing tasks and step up only when the composition demands it. High complexity can push generation time past two minutes on difficult prompts. There is no additional cost for any level — the tradeoff is purely latency. The complexity parameter documentation breaks down the planning-depth mechanics in detail.
The Catch Before You Migrate: Batch Support Is Gone
This is the item that did not make the headline. GPT-Image-2.5 supports the Batch API, which runs asynchronous jobs at half the standard rate — $15 per million output tokens instead of $30. GPT-Image-2.6 does not support Batch.
If you have a nightly image generation pipeline, a bulk asset-creation job, or any workflow that tolerates async latency in exchange for cost savings, migrating to 2.6 today doubles your image output costs. The workaround is to stay on 2.5 for those pipelines until OpenAI adds Batch support to 2.6. This post on the Batch API removal details the cost impact and migration considerations.
For everything else, migration is a one-line change: update the model ID from gpt-image-2.5-flare to gpt-image-2.6-flare (or -sunburst). The endpoints and parameters are otherwise identical. The full OpenAI image generation API reference documents every parameter for both models.
