Source: https://docs.sogni.ai/models/gpt-image-2-6/

# GPT Image 2.6 Sunburst and Flare

Select `gpt-image-2.6-sunburst` or `gpt-image-2.6-flare` explicitly. Both are partner models billed in Premium Spark on the Fast network. Sunburst targets detailed work; Flare targets faster iteration. Existing GPT Image selectors keep their existing behavior.

## [#](https://docs.sogni.ai/models/gpt-image-2-6/#animated-gifs)Animated GIFs

GIFs are text-to-image generations. Describe the motion and how the last frame should return to the first. References, masks, annotation layers, transparent backgrounds, and output compression cannot be combined with GIF output.

```
{
  "model": "gpt-image-2.6-flare",
  "prompt": "A small clay monkey gently lifts a glowing lantern, blinks, and lowers it to its starting position. Locked camera, seamless loop.",
  "width": 896,
  "height": 896,
  "gptImageQuality": "medium",
  "gptImageComplexity": "high",
  "outputFormat": "gif",
  "gptImageFrameCount": 9,
  "gptImageFrameDurationMs": 200
}
```

Use this argument object with `generate_image` through the [direct generation API](https://docs.sogni.ai/api-reference/direct-generation/). The JavaScript SDK uses `gptImageFrameCount`, `gptImageFrameDurationMs`, and `gptImageComplexity` on `projects.create()` and `projects.estimateCost()`; Python uses their snake\_case equivalents.

| Setting | Supported values |
| --- | --- |
| Frames | 2–16; default 9 |
| Time per frame | 20–2000 ms; rounded by OpenAI to 10 ms; default 200 |
| Frame dimensions | Positive multiples of 16; maximum 3:1 aspect ratio; must fit a supported frame sheet below |
| Frame sheet | 655,360–8,294,400 pixels, including unused cells; maximum 3840 per edge and 3:1 aspect ratio |
| Default frame size | 896 × 896 for 9 frames |

Sogni validates frame sizes against these layouts (verified November 4, 2026). Multiply the frame width and height by the sheet columns and rows; at least one orientation must meet the sheet limits above. Counting only finished frame pixels misses unused cells and can admit an invalid request.

| Frame count | Sheet columns × rows |
| --- | --- |
| 2–8 | 4 × 2 or 2 × 4 |
| 9 | 3 × 3 |
| 10–12 | 4 × 3 or 3 × 4 |
| 13–16 | 4 × 4 |

The JavaScript SDK exports `resolveGptImageGifSize()` to infer valid dimensions and `getGptImageGifSheet()` to validate explicit dimensions. For example, 1280 × 1280 with five frames is invalid because both sheet orientations exceed 3840 pixels per edge. Tiny frames can also fall below the minimum sheet area.

Nine frames at 200 ms produce a 1.8-second animation. Changing frame timing changes playback speed. Smaller frames are supported, including 320 × 320. As of November 4, 2026, OpenAI's live API enforces 16 frames despite its guide describing a higher limit.

PNG and WebP output are still images. Transparent PNG and WebP stills are supported; transparent animated PNG (APNG) and transparent GIF generation are not. Download a GIF using its `.gif` extension to preserve animation.

## [#](https://docs.sogni.ai/models/gpt-image-2-6/#planning-complexity)Planning complexity

`gptImageComplexity` accepts `low`, `medium`, or `high` (default). It controls planning independently of the rendering quality setting. Rendering quality remains `low`, `medium`, `high`, `xhigh`, or `max`. Specify concrete quality when requesting a quote.

## [#](https://docs.sogni.ai/models/gpt-image-2-6/#annotation-guided-edits)Annotation-guided edits

With `edit_image`, pass `markup_image_url` alongside a reference image and an edit prompt. The annotation is a separate transparent PNG whose dimensions match the first reference after orientation is applied. Use arrows, circles, or labels to indicate the requested change. The source and annotation must each fit 16,777,216 pixels; the annotation must be under 50 MB. Do not combine markup with an edit mask.

On SDK projects, the annotation field is `gptImageMarkupUrl` (Python: `gpt_image_markup_url`). References remain ordered; the annotation applies to the first reference. PNG/WebP transparency, JPEG/WebP compression, masks, and up to 16 references remain available for still-image requests subject to the selected model's limits.

## [#](https://docs.sogni.ai/models/gpt-image-2-6/#pricing-and-observed-performance)Pricing and observed performance

As of November 4, 2026, OpenAI publishes the same per-token rates for 2.5 and 2.6: $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens. At 1024 × 1024 Medium, our requests returned 439 output tokens on 2.5 and 1,449 on 2.6: $0.01317 and $0.04347 in provider output cost. Prompt and reference input add cost. Sogni applies its partner-model margin and quotes in Premium Spark.

GIFs reserve the most expensive supported sheet orientation, including unused cells. The quote reserves a conservative output budget, and the final charge is reduced to confirmed usage when lower. Use the quote for the exact frame size, count, quality, and references you intend to send.

Our small direct-provider comparison used three prompts per model at 1024 × 1024, Medium, JPEG, one take per prompt. Median response time was 6.55 seconds for 2.6 Flare versus 10.48 for 2.5 Flare, and 19.99 seconds for 2.6 Sunburst versus 14.96 for 2.5 Sunburst. These measurements exclude Sogni delivery and do not establish a general speed or quality ranking.

See the [GPT Image 2.6 versus 2.5 comparison and GIF examples](https://blog.sogni.ai/blogs/gpt-image-2-6/), [OpenAI's image guide](https://developers.openai.com/api/docs/guides/image-generation), [pricing](https://developers.openai.com/api/docs/pricing), and the [Sogni model page](https://www.sogni.ai/models/gpt-image-2-6).