GPT Image 2.6 vs 2.5: create GIF animations

What changes from GPT Image 2.5?

On Sogni	GPT Image 2.5	GPT Image 2.6
Animated output	Still images	GIF animations from text
Planning complexity	No separate control	Low, Medium, High
Rendering quality	Low through Maximum	Low through Maximum
Transparent still images	PNG and WebP	PNG and WebP
Image edits and references	Supported	Supported; separate annotation layer available

Choose 2.6 when the idea needs motion, or when you want to control planning separately from rendering. GPT Image 2.5 remains a useful choice for still images, particularly at Medium where its measured output cost is lower. Both generations keep Sunburst and Flare choices.

Ask for the motion and the loop

Choose GPT Image 2.6 Sunburst or Flare and GIF output. Describe a short action with a clear return to its starting pose. The default nine frames at 200 milliseconds each give you a 1.8-second animation. A locked camera helps keep the loop readable.

A monkey in a green coat lifts a glowing lantern, blinks, and lowers it again in front of a snowy cabin. A seamless loop: the monkey returns exactly to its starting pose after 1.8 seconds. Fixed camera, crisp detailed illustration.

The Sunburst example above finished in 34.13 seconds. A similar Flare prompt took 24.59 seconds at the same size and frame settings. Each used 4,386 image output tokens, about $0.132 in provider cost including its text input. The corresponding Sogni charge at the standard partner margin is about $0.172, or 34.3 Premium Spark. The live quote accounts for the request you actually send.

GIFs are generated from text. References, edit masks, annotation layers, transparent backgrounds, and compression settings do not combine with GIF output. You can choose 2–16 frames and 20–2000 milliseconds per frame, with frame-size limits that depend on the frame count. Nine 896 × 896 frames are the default; increasing the count can require smaller frames. The API guide explains the size limits. More frames give a short action room to develop: our two- and four-frame animations raised the lantern without completing the requested return, while eight and sixteen frames completed it.

Sunburst GIF of a monkey raising and lowering a lantern — Sunburst · 9 frames · 200 ms/frame · Medium
Download Sunburst GIF

Flare GIF of a monkey with a glowing lantern — Flare · 9 frames · 200 ms/frame · Medium
Download Flare GIF

Download the animation as a .gif file to keep its motion. PNG and WebP output are still images; transparent animated PNG (APNG) output is not supported. GIF generation uses an opaque background.

Two more controls for still images

Planning complexity is separate from rendering quality. Low, Medium, and High control planning; High is the default. You still choose Low, Medium, High, Extra high, or Maximum rendering quality.

Annotation-guided edits use a separate transparent PNG over the first reference image. Circle the vase you want recolored or draw an arrow toward the area to change, then describe the change in the prompt. The annotation and first reference need matching dimensions. An annotation cannot be combined with an edit mask.

Same token rates, different output budgets

OpenAI lists the same 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. A quality label can still cost a different amount because the models use different output budgets.

1024 × 1024 Medium	2.5	2.6
Measured output tokens	439	1,449
Provider output cost	$0.01317	$0.04347
Sogni output cost, standard margin	$0.01712	$0.05651

These output-only figures exclude prompt and reference input. At High, Extra high, and Maximum, the models share the output-budget formula for still images. GIF quotes reserve a conservative amount because frame packing can vary; final charges decrease to confirmed usage when lower.

Flare was quicker in our small sample; Sunburst took longer

We sent three identical prompts to each model: a lettered bakery poster, a labeled solar-system board, and a watch photograph. Every request used 1024 × 1024, Medium quality, JPEG output, and one take. The 2.6 requests used High planning complexity. Calls ran serially through OpenAI's Images API on November 4, 2026 PT. These are provider-response times and exclude Sogni queue and delivery time.

Median response time, three prompts	2.5	2.6
Flare	10.48 s	6.55 s
Sunburst	14.96 s	19.99 s

All twelve images had readable requested text. Composition and detail differed, and neither generation won every comparison. The watch images still needed inspection for physical details. Three prompts and one take each are too small a sample to establish a general speed or quality ranking.