GPT Image 2.6 Flare vs Sunburst — SVG and GIF output, real costs in ₹, and when to switch

Key takeaways

• OpenAI released GPT Image 2.6 on 4 November 2026 as two API models: gpt-image-2.6-flare (fast) and gpt-image-2.6-sunburst (more detail, steadier edits) (OpenAI Developer Community).
• The big change is the output format, not the picture quality. Both models can return a vector SVG or an animated GIF, as well as PNG, JPEG and WebP (OpenAI image guide).
• The token price is unchanged at US$30 per million image-output tokens, but each image may use more tokens. One developer measured 1,136 output tokens for a medium image on 2.6 against 344 on 2.5. That is roughly ₹3.2 per image instead of ₹1, a single sample, not an official figure.
• There is no Batch API for 2.6 yet, so the 50% batch discount you may use on 2.5 does not apply (DEV Community).
• Our advice: try 2.6 for icons, illustrations and short GIFs, where SVG and animation save real design hours. Keep high-volume product-photo pipelines on 2.5 until you have measured your own cost per image.

1. What shipped on 4 November

OpenAI's announcement pitches Images 2.6 around three things: animated GIFs from a prompt or an existing image, saved reference images (people, pets, characters) that you can reuse across generations, and faster output. OpenAI says generation is about 50% faster and that Flare can finish an image in "as little as six seconds", a best case at low quality (OpenAI Developer Community, Hugging Face guide).

For developers, the API changes matter more:

• output_format: "svg" on the generations endpoint returns a static vector file with no scripts, external links or embedded bitmaps. It supports transparent backgrounds.
• output_format: "gif" with frame_count and frame_duration_ms. OpenAI's own example uses 9 frames at 200 ms, a 1.8-second loop.
• Markup-guided edits: you draw on a transparent PNG overlay (circle the area, scribble a note) instead of building a precise mask.
• A new complexity setting (low / medium / high, default high) that controls how much the model plans the layout, separate from quality, which controls detail.

Sizes are auto or any custom size in multiples of 16, with aspect ratios from 1:3 to 3:1 (OpenAI image guide).

2. Flare vs Sunburst

Prices and endpoint support are as reported from OpenAI's pricing and model pages (DEV Community, LLM Gateway). The two models cost the same per token, so the choice is about speed against fidelity, not price.

A simple rule: use Flare while you are exploring (mood boards, ten banner variants, a quick social post) and switch to Sunburst for the final edit, especially when a face, a logo or a product has to stay identical across several changes.

3. What it really costs per image in ₹

OpenAI bills image output by tokens, not a flat fee per image, and it has not published per-image figures for 2.6. The only public data point so far is from one developer on OpenAI's forum: the same prompt at medium quality used 344 output tokens on 2.5 and 1,136 on 2.6 (OpenAI Developer Community). The same person estimated that low and medium quality cost roughly 3–4× more on 2.6, while high quality and above cost about the same.

At about ₹94 per US$ and US$30 per million output tokens, that sample works out like this:

Budget 18% GST on top: imported online services attract IGST (paid under reverse charge if your company is GST-registered). Input tokens (your prompt and any reference images) are extra but usually small next to output.

Treat the 2.6 column as a starting estimate. Run 50 of your own real prompts on both versions and read the usage field in each API response before you change a production pipeline.

4. Where SVG and GIF output actually help

SVG is the most useful part of this release for web teams. An SVG icon or illustration is usually a few kilobytes, scales cleanly from a 16 px favicon to a hero banner, and can be recoloured in CSS. A PNG cannot do any of that. For Indian agencies and in-house teams that make dozens of feature icons, empty-state illustrations or simple infographics for every client site, this can replace hours of tracing in Figma or Illustrator.

Know the limits before you plan around it:

• One SVG per request (n=1), with no streaming or partial previews.
• You cannot edit an existing SVG through the edits endpoint. Edit a bitmap, then generate a new SVG.
• SVG billing has two parts: the image model renders a reference image first, then an authoring model writes the vector code, and you pay for both.
• Always open the file and check it. Clean up path counts with a tool like SVGO before shipping, and do not assume complex scenes vectorise well.

GIFs suit WhatsApp Business catalogues, email headers, onboarding hints and social posts, where a short loop gets more attention than a still. Keep loops short and small: a 512×512, 9-frame GIF is fine for chat, but a long, large GIF will hurt page weight and your Core Web Vitals. For a website, convert it to a muted, looping MP4 or WebM instead.

5. Risks and gaps to plan for

• No dated snapshot. The 2.6 model pages list only the alias, so OpenAI can change the model behind the same name. For production, record the exact prompt and settings, and re-test after any announced update.
• No Batch API. Overnight jobs that relied on the 50% batch discount will cost twice as much per token if you move them now.
• Edit drift. OpenAI's own guide warns that repeated edits can slowly change details that should stay fixed. For product shots, paste the approved edit back onto the original image rather than chaining edits.
• Reference images of real people. Saved references of faces make consistent characters easy, which also makes misuse easy. Get written consent before using a customer's or employee's photo, and remember that India's DPDP Act treats a face photo as personal data.
• Codex availability. OpenAI says the GIF features are rolling out in ChatGPT and Codex, but one developer reported on 6 November that it was not yet live in Codex. Check before you build a workflow on it.

6. What to do this week

1. Designers and agencies: generate your next icon set as SVG with Flare, run it through SVGO, and compare the time spent with your usual process.
2. E-commerce teams: do not switch your bulk product-image job yet. Run a 50-image A/B test of 2.5 against 2.6 Sunburst and compare cost and quality side by side.
3. Developers: put the model name in config, not code, log usage tokens per request, and add a cost alert, so a 3× jump shows up on day one and not on the invoice.
4. Founders: for marketing GIFs and illustrations, the ChatGPT app may be cheaper than the API at low volume. Use the API only when you need automation.

Want image generation wired into your product or CMS with cost controls? See our AI & automation development service, or shrink what you ship with our free image compressor.