GPT Image 2.6 is out: what changed and what it costs

On 4 November 2026 OpenAI released GPT Image 2.6, the next generation of its image model. In the API it is gpt-image-2.6-flare and gpt-image-2.6-sunburst; on our platform the public id is gpt-image-2-6. Here is what actually changed since 2.5, how the two variants differ, what it costs and how to make your first call.

Official sources: the OpenAI API changelog of 4 November 2026, and the model pages for Flare and Sunburst.

What is new compared with 2.5

The release is not just "prettier images". OpenAI lists three new capabilities.

• Complexity control. A new complexity field takes low, medium or high (default high) and decides how much the model plans before rendering. Lower complexity suits focused edits such as changing a background or a sign; higher complexity helps with dense text, complex layouts and relationships between objects. There is no separate fee, but higher complexity can add latency.
• SVG and animated GIF output. OpenAI added output_format: "svg" for editable vector graphics (icons, illustrations, logos) and an animated GIF with a configurable frame count and duration. Both are generation-only, not for edits.

The changelog also notes that markup-guided edits now work with both 2.5 and 2.6: you attach an annotation layer with the area circled and describe the change in the prompt.

What is live on our side. The complexity control works on both of our routes — complexity reaches the model. SVG, animated GIF and markup-guided edits are not enabled yet: our routes return PNG and do not accept an annotation layer. We will not advertise them until each is proven on a live job; they are listed here as capabilities of the model itself.

Flare and Sunburst: which one to pick

They are not "old and new" versions but two characters of the same model:

• Flare is the fast variant for everyday work and iteration. OpenAI describes it as built for quick iteration with stronger layouts for posters, product imagery and other professional visuals.
• Sunburst is built for detailed work: tighter composition and more consistent successive edits, so the same subject stays recognisable across several rounds of revision.

On our platform you choose the variant with tier: flare (default) or sunburst. Both accept the same parameters, so comparing them on your own prompt is a one-field change.

What it costs

Prices are struck on the full, non-promotional supplier cost with a 35 % gross margin and rounded up, so rounding can never eat into the margin. You pay only for a successful generation: a failed job is not charged.

The live per-image price is on the model page — it is read from the catalog rather than typed into this article, so it cannot go stale. Two rules explain how it moves:

• Resolution sets the lane. 1K is the cheapest cell; 2K and 4K cost more because they come from a different supplier.
• Quality moves the price at 2K and 4K only. At 1K the image comes from a supplier with one flat per-image price, while larger tiers come from a supplier that charges a flat price per output image by quality. Within one resolution and quality the price does not depend on the aspect ratio or on how many reference images you send.

Copy-paste example

Through the OpenAI-compatible endpoint:

curl https://getvibeapi.com/v1/images/generations \
  -H "Authorization: Bearer sk-pv-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2-6",
    "prompt": "Poster for a ceramics exhibition: three matte vases on a warm sand background, soft side light, close-up, room for a headline at the top",
    "size": "1536x1024",
    "quality": "high",
    "extra_body": { "resolution": "2K", "tier": "sunburst", "complexity": "high" }
  }'

Through the native endpoint — the same request, with the options in opts:

curl https://getvibeapi.com/api/media/generate \
  -H "Authorization: Bearer sk-pv-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2-6",
    "prompt": "Poster for a ceramics exhibition: three matte vases on a warm sand background, soft side light, close-up, room for a headline at the top",
    "opts": { "resolution": "2K", "quality": "high", "tier": "sunburst", "complexity": "high", "aspect_ratio": "3:2" }
  }'

/api/media/generate answers with a jobId; collect the result from GET /api/media/jobs/<jobId>. The /v1/images/generations address waits for the finished image and answers in the OpenAI shape.

Parameter notes

• Aspect ratio. 1K accepts the familiar ratios, including 21:9. At 2K and 4K the set is narrower: 21:9 is unavailable there, and a request that asks for it is refused before any charge rather than quietly replaced with the nearest size.
• Editing. Send a source image (image_url or input_urls) and describe the change in prompt. For markup-guided edits attach the annotation layer.
• Format. PNG. The OpenAI-compatible adapter does not accept output_format, and SVG/GIF are not enabled on our routes yet — such a request is refused rather than quietly replaced.

Who should use it

GPT Image 2.6 earns its price where layouts and text in the image matter: posters, covers, product cards, ad creatives. For quick drafts use Flare with quality: "low" — the cheapest cell. For final assets use Sunburst with high, and raise complexity only when the frame really is dense.

If you need the previous character of generation, GPT Image 2.5 and GPT Image 2 stay in the catalog at their own prices — there is no need to change existing integrations.