OpenAI Images 2.6: GIF Creation, Flare vs. Sunburst, API Pricing, and the Fine Print
OpenAI has shipped Images 2.6, and the headline is motion. ChatGPT, Codex, and API users can now generate GIFs from a text prompt or a rough sketch, reuse saved reference images across chats, and export SVG files—while OpenAI claims image generation is 50% faster than before[reference:0].
The release also splits the API into two models: GPT-Image-2.6 Flare for speed and GPT-Image-2.6 Sunburst for detail and editing consistency. Below is what actually shipped, what it costs, and what early testers are already flagging.
What Images 2.6 Actually Adds
OpenAI says people now create about 4 billion images per week across ChatGPT, Codex, and the API—more than half a billion per day[reference:1]. Images 2.6 is available globally in ChatGPT and Codex across all subscription tiers on web, desktop, and mobile, including free[reference:2].
The release is not just a speed bump. The three features that matter most for creators and developers are:
- Native GIF creation from prompts, sketches, or reference images.
- Reusable reference images that can be saved and recalled by name.
- New API controls: image markup, SVG export, and a complexity parameter.
GIF Creation: Promising, But Not a Production Tool Yet
ChatGPT can now generate a GIF in a single step from a prompt such as “can you make a gif of a cat dancing?” or from an uploaded photo[reference:3]. Once the GIF is generated, users can adjust playback speed, inspect individual frames, and share the result [Tweet].
OpenAI positions GIFs for marketers, creators, and small teams: animating a brand mascot for a campaign, making reactions and memes for social media, or using animated storyboards to explore character movement before video production[reference:5]. A rough sketch can also become an animated character—no polished design file required[reference:6].
But there are real limits. In the API, GIF output supports 2 to 64 frames, with each frame lasting 20 to 2,000 milliseconds (default 9 frames at 200ms each). GIF output is available through the generations endpoint, not the edit endpoint[reference:7]. A hands-on test found that motion and subject consistency are “generally usable,” but fine details shift, hand relationships break, and frame counts or sizes do not always follow the prompt exactly[reference:8].
Practical takeaway: GIFs are good for reactions, memes, mascots, and storyboard drafts. For brand-accurate product animation or anything requiring frame-perfect continuity, plan on manual review and frame-level fixes.
Reusable Reference Images: Save Once, Call by Name
Images 2.6 lets ChatGPT users save a reference image and reuse it in later conversations without re-uploading. OpenAI’s example: save a photo of your dog as “Pepper,” then ask for “a watercolor portrait of @Pepper”[reference:9]. The same mechanism works for characters, pets, products, and brand mascots[reference:10].
The goal is continuity. Instead of rebuilding context every session, the model carries the same subject forward. Early tests suggest this helps with identity drift, but it is not a guarantee—when likeness shifts, the practical fix is to constrain the prompt further and attach the reference again[reference:11].
Flare vs. Sunburst: Which API Model Should You Use?
OpenAI has split the API into two models with the same standard token rates but different priorities[reference:12].
| Model | Best for | What OpenAI says | Caveat |
|---|---|---|---|
| GPT-Image-2.6 Flare | Fast iteration, social drafts, poster concepts, product variations | Can generate at low quality in as little as 6 seconds; improves layout quality for posters and product imagery[reference:13] | Early testers report transparency-grid artifacts, broken lines, and over-simplified designs in some outputs[reference:14] |
| GPT-Image-2.6 Sunburst | Final assets, brand consistency, character sheets, multi-round edits | Higher quality and greater consistency across edits at medium or high quality[reference:15] | Slower than Flare; OpenAI does not publish an absolute latency figure for Sunburst[reference:16] |
The practical workflow most testers recommend is Flare for drafts, Sunburst for finals. Flare is the iteration engine—use it when you will throw away half the outputs anyway. Sunburst is the consistency engine—use it when a logo, product, or character must survive a chain of edits[reference:17].
One important note: OpenAI does not describe a ChatGPT variant selector. In ChatGPT, you get Images 2.6 without choosing Flare or Sunburst. API users must set the model explicitly to gpt-image-2.6-flare or gpt-image-2.6-sunburst[reference:18].
API Pricing: Same Rates, But Watch Token Usage
Both Flare and Sunburst share the same standard token pricing:
- Text input: $5.00 per 1M tokens (cached: $1.25)
- Image input: $8.00 per 1M tokens (cached: $2.00)
- Image output: $30.00 per 1M tokens[reference:19]
OpenAI provides a calculator for per-image estimates. At the “max” token budget, Sunburst output works out to roughly $0.21 per image[reference:20]. That is not the final word, though: independent analysis suggests that at 1024×1024 in low and medium quality, Images 2.6 may consume about four times the tokens of Images 2.5, which could make the effective per-image cost higher even though the sticker rates are unchanged[reference:21].
Cost check: Run your own prompt and quality settings through OpenAI’s calculator before committing to a production budget. Token-based pricing means the same dollar rate can produce very different per-image costs depending on complexity, quality, and output format.
New API Controls: Markup, SVG, and Complexity
Images 2.6 brings several controls that matter for professional workflows:
- Image markup in the API: The markup feature already available in ChatGPT now works through the API, so teams can pinpoint exactly where a change should happen[reference:22]. You send a transparent PNG overlay with circles or sketches marking the area, plus a prompt[reference:23].
- SVG export: Images 2.6 can export SVG files for websites, apps, and print—a meaningful step beyond raster output for design handoff[reference:24][reference:25].
- Complexity parameter: A new setting controls how much planning the model does before rendering. Lower complexity suits quick edits; higher complexity helps with dense text, complex layouts, and scenes with many constraints. Complexity is separate from quality[reference:26].
Benchmarks and Independent Testing: Speed Is Real, Artifacts Are Too
OpenAI’s 50% speed claim is self-reported, and independent benchmark rankings were not yet available at launch[reference:27]. Early Arena.ai data gives a mixed but useful picture:
- In Text-to-Image Arena, Sunburst (Medium) ranked #2 with 1,409 points; Flare (Medium) ranked #4 with 1,384 points[reference:28].
- Sunburst landed #1 in Cartoon, Anime & Fantasy[reference:29].
- In Image Edit Arena, Sunburst (Medium) ranked #3; Flare (Medium) ranked #5[reference:30].
Quality reports are not uniformly positive. One tester rolled back to 2.5-Flare after 20 minutes, citing “too many artifacts (transparency grids baked in, broken lines) as well as the designs being overly simplistic”[reference:31]. That is one data point, not a benchmark, but it names a failure mode OpenAI does not quantify.
Availability and Model IDs
Images 2.6 is available today in ChatGPT and Codex globally across all subscription tiers on web, desktop, and mobile[reference:32]. In the API, the model IDs are:
gpt-image-2.6-flaregpt-image-2.6-sunburst
Both support complexity control, markup-guided edits, SVG output, and animated GIF output through the Image API[reference:33].
The Bottom Line
Images 2.6 is a meaningful update, but it is not a “set it and forget it” creative engine. GIFs are the flashiest addition and genuinely useful for reactions, mascots, and storyboard drafts—yet frame-level inconsistencies and artifact reports mean brand-critical animation still needs human review.
For developers, the Flare–Sunburst split is the smarter part of the release. Use Flare for speed and iteration, switch to Sunburst for final assets and multi-edit consistency, and model your token spend carefully before scaling. The new markup, SVG, and complexity controls are the quiet upgrades that may end up mattering more than the GIF headline.