GPT Image 2.6: Flare vs Sunburst, what's new, pricing and API

GPT Image 2.6 at a glance

What's new in GPT Image 2.6

OpenAI's announcement and developer documentation list these changes compared with Images 2.5:

OpenAI also said its products now generate 4 billion images a week, up from 3 billion two months earlier.

GPT Image 2.6 Flare vs Sunburst

Both models take the same inputs, support the same settings and formats, and cost the same per token. They differ in what OpenAI tuned them for, and in speed:

Same prompts and edits run on both models: GPT Image 2.6 Flare vs Sunburst, with side-by-side tests.

GPT Image 2.6 examples from GPT Image Hub

We made these on GPT Image Hub on 5 November 2026 with GPT Image 2.6 Flare Saver and GPT Image 2.6 Sunburst Saver, at 10 credits each. Times run from request to finished image through the Hub's Saver route and were slower than the launch-night API test above: 19–21 seconds for Flare and 17–38 seconds for Sunburst. Images requested at 1024×1024 came back at 1254×1254.

Posters and infographics

Product shots and illustrated maps

Edits that keep the rest of the image

More outputs, prompts and timings are in our GPT Image 2.6 Flare Saver guide and GPT Image 2.6 Sunburst Saver guide.

GPT Image 2.6 vs GPT Image 2.5

GPT Image 2.5 (8 September 2026) also came as Flare and Sunburst. The figures below are from CellCog's launch-night test (80 generations on 5 November 2026, one machine, wall time from request to bytes), using the same prompt, size and quality on both versions; token counts repeated exactly across reruns, so they are the most reliable numbers.

• From high quality upward, 2.6 Flare costs exactly the same as 2.5 Flare and was 2.1–2.5 times faster in the test, so it is a free upgrade for final-quality work.
• At low and medium quality, 2.6 bills 3.3–4.2 times more tokens than 2.5. For high-volume drafts or thumbnails, 2.5 is still much cheaper, and the test found the quality difference hard to see on small assets.
• 2.6 has no Batch pricing, so overnight bulk jobs on 2.5 Batch cost about a quarter of the same work on 2.6.
• 2.6 Sunburst was no faster than 2.5 Sunburst; it preserved a detail slightly better in a three-step edit chain, at higher cost.

Read our measured guides to GPT Image 2.5 Flare and GPT Image 2.5 Sunburst, or the full GPT Image 2.6 vs 2.5 comparison with the same prompts on both versions.

How much does GPT Image 2.6 cost?

OpenAI bills GPT Image 2.6 by token, not by image, and Flare and Sunburst share one price list. What you pay for an image depends on how many image output tokens it produces, which grows with quality and size:

• 1024x1024 at low quality: 824 output tokens, about $0.025.
• 1024x1024 at medium quality: 1,449 output tokens, about $0.044.
• 1024x1024 at high quality: 1,758 output tokens, about $0.053.
• 1024x1024 at max quality: 7,024 output tokens, about $0.21.
• A 12-frame animated GIF at medium quality cost $0.056 in the test; a 16-frame GIF at high quality cost $0.108.

Token counts are from CellCog's test; larger sizes use more tokens. SVG requests are billed twice: the reference image at the image model's rate, plus the SVG authoring tokens. The complexity setting has no separate fee.

On GPT Image Hub, GPT Image 2.6 Flare Saver and GPT Image 2.6 Sunburst Saver cost a flat 10 credits per image (about $0.010–$0.013, depending on the credit pack) whatever the size and quality, so a 4K image at max quality costs the same as a low-quality draft.

The new features in detail

Using GPT Image 2.6 in the API

Both models use the Image API you already use for GPT Image 2 and 2.5; switching is a one-line model change. A standard generation, with complexity lowered for a simple scene:

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-flare",
    "prompt": "A matte black wireless speaker on a pale concrete plinth, soft window light, blank space on the right for a headline",
    "size": "1536x1024",
    "quality": "high",
    "complexity": "low"
  }'

An animated GIF with Sunburst: 9 frames at 200 ms, a 1.8-second loop.

{
  "model": "gpt-image-2.6-sunburst",
  "prompt": "A paper lantern swaying gently in the wind at night, warm glow, the lantern returns to its starting position at the end",
  "output_format": "gif",
  "frame_count": 9,
  "frame_duration_ms": 200,
  "size": "1024x1024",
  "quality": "medium"
}

• Organization verification is required before you can call GPT Image models.
• partial_images (0–3) streams previews of PNG, JPEG and WebP results, but not of GIFs or SVGs.
• Prefer jpeg output when latency matters; use png or webp with background set to transparent for cutouts.
• For end-user apps, pass a hashed safety_identifier (128 characters or fewer) instead of the deprecated user field.

How to prompt GPT Image 2.6

• Keep the prompt structure that worked on GPT Image 2 and 2.5: subject, setting, composition, lighting, style, then any on-image text in quotes.
• Put exact text in quotation marks and keep it short. In CellCog's poster test both 2.5 and 2.6 spelled every line correctly.
• Set complexity to low for one-change edits and leave it at high for infographics, menus and scenes with many objects.
• For edits, say what to keep and what to change; for local changes, mark the area with markup instead of describing its position.
• For GIFs, describe one small, repeating motion that ends where it starts, and keep the frame count low.
• For SVG, ask for flat shapes, a limited palette and no gradients, photos or texture.

Six prompts written for the new features. The poster, product and infographic prompts work as written in ChatGPT, the API or the GPT Image Hub generator; the GIF, SVG and markup prompts need the API's output_format and markup fields (ChatGPT also makes GIFs).

For the full method, with tested prompts next to their results, see how to prompt GPT Image 2.6.

Where to use GPT Image 2.6

Sources

Specifications are from OpenAI's documentation as of 5 November 2026. Timings and token counts are from CellCog's independent test, not OpenAI benchmarks; your results will vary with size, quality and load.

• OpenAI: GPT-Image-2.6 Flare model page
• OpenAI: GPT-Image-2.6 Sunburst model page
• OpenAI image generation guide
• OpenAI on X: introducing ChatGPT Images 2.6 (4 Nov 2026)
• OpenAI on X: availability in ChatGPT, Codex and the API
• TestingCatalog: OpenAI launches Images 2.6 with GIFs and image reference
• CellCog: GPT Image 2.6 vs 2.5, 80 generations on launch night