GPT Image 2.6 Sunburst vs Flare: Same Price, 2.3x Faster

GPT Image 2.6 Flare is the model to use for banners and social images with text. In our API test it rendered every requested line of text exactly, cost the same $0.034 per 1536×1024 image as Sunburst, and returned in about 9 seconds against Sunburst's 21. Sunburst gave us nothing extra for the wait except a header line we never asked for.

OpenAI released Images 2.6 on 4 November 2026 as two API models: Sunburst, described as being for "detailed visual work", and Flare, for "fast iteration". We sent both the same text-heavy banner prompt over the API and timed each call, priced each result from the usage field, and checked the new SVG and GIF outputs.

At a glance

Key results

• Same price per image. Every 1536×1024, medium-quality banner used 1,132 image output tokens on both models: $0.0343–$0.0344 at the $30 per million output-token list price.
• Flare is about 2.3 times faster. Sunburst took 20.0–21.2 seconds per image over three calls; Flare took 8.0–9.2 seconds.
• Text came out right on both. All six banners spelled every requested line exactly, including a Vietnamese headline full of accents ("Khuyến mãi cuối năm: Giảm 30% mọi khoá học Excel").
• Sunburst adds things. In all three of its banners it invented a small "Office Training" header with a tagline. Flare stuck to the three lines we asked for.
• SVG output is editable, but the layout broke. The SVG kept the text as real <text> elements; in our render the headline ran off the right edge and the date line hid behind the button.
• GIF works out of the box. A 4-frame animated GIF came back in 8.3 seconds for $0.0169.

Hands-on test

Setup

• Date: 5 November 2026, from a Python script (httpx) on a Windows 11 PC in Vietnam (WSL, Ubuntu 24.04).
• Access: the project's own OpenAI API key, POST /v1/images/generations. We have no paid ChatGPT plan, so this tests the API only, not Images 2.6 inside ChatGPT.
• Models: gpt-image-2.6-sunburst and gpt-image-2.6-flare, size: 1536x1024, quality: medium, default complexity (the guide says both default to high).
• Prompt: a flat web banner for an office training company with three exact text lines (headline, an end-date line, a button label), navy background, "simple spreadsheet icon". One English and one Vietnamese version.
• Runs: English once per model, Vietnamese twice per model. Cost is computed from the usage object each response returned, at the prices on OpenAI's model pages.

Text in the image: Sunburst vs Flare

Both models put the headline, the "Offer ends 31 December" line and the "Enroll now" button on the banner with no spelling mistakes. The difference was discipline: Sunburst added an "Office Training Company" logo block on its own initiative every time, which is the kind of change you would have to edit out of a real ad. OpenAI's guide warns that higher complexity "can increase latency and introduce unwanted changes", and that is what we saw from the slower model at the default setting.

One thing both models got wrong in the same way: we asked for a "simple spreadsheet icon" and every banner came back with a green icon that looks very close to the Microsoft Excel logo. If you plan to run these as ads, ask for a generic icon explicitly or replace it, because a lookalike trademark is a risk you add without asking for it.

Non-English text: a Vietnamese test

We repeated the prompt with Vietnamese text, which stacks accents on top of accents ("Khuyến", "Ưu đãi", "Đăng ký"). All four Vietnamese banners rendered every diacritic correctly, with timings that matched the English run (Sunburst 20.0 and 21.2 s, Flare 8.0 and 9.2 s). OpenAI's documentation does not make any claim about non-Latin text; the guide still notes that the model "can still struggle with precise text placement and clarity", so check dense small print before publishing.

SVG and GIF output

SVG. Setting output_format: "svg" failed at first with "SVG generation requires 'svg.model'"; the guide does not document that field. Passing an authoring model, "svg": {"model": "gpt-5-mini"}, worked. The call took 30.3 seconds. The file was 4.6 KB of clean XML with the text as editable <text> and <tspan> elements set in Arial, which is the useful part: you can fix the wording without regenerating. But the layout depends on the fonts of whatever renders it, and in our render the headline and the date line overlapped other elements. The response's usage only listed the 1,132 image tokens ($0.0346); the guide says SVG also bills the authoring model's tokens, and those were not itemised in the response.

GIF. Our first try with 4 frames at 1024×1024 was rejected because the frames did not fit the model's sheet size limit. With size: "auto" the API picked 512×512 and returned a 4-frame loop in 8.3 seconds for 555 output tokens, $0.0169.

What worked

What worked

• Exact text, including Vietnamese diacritics, on 6 of 6 banners.
• Flare at the same price as Sunburst and less than half the wait.
• GIF output from a single API call, cheaper than one still banner at our sizes.

Where it fell short

What didn't work

• Sunburst adding an unrequested header on 3 of 3 runs.
• An Excel-lookalike logo on every banner from a "simple spreadsheet icon" prompt.
• SVG: an undocumented required parameter, a layout that overflowed when rendered, and an incomplete cost figure in usage.

Our verdict

Use Flare for anything with text: banners, thumbnails, social cards. It was as accurate as Sunburst on our prompt, cost exactly the same and came back in under 10 seconds. We would only reach for Sunburst after Flare misses on a specific image, and then compare complexity settings rather than assume the slower model is better. Skip SVG output for finished designs for now; treat it as an editable starting point. If you would rather pay a flat monthly plan than per image, see what Higgsfield's plans cost.

FAQ

What is the difference between GPT Image 2.6 Sunburst and Flare?

OpenAI positions Sunburst for detailed visual work and Flare for fast iteration. Both have the same prices and endpoints. In our test Flare was about 2.3 times faster and Sunburst added unrequested elements.

How much does a GPT Image 2.6 image cost?

Both models charge $30 per million image output tokens. A 1536×1024 medium-quality image used 1,132 output tokens in our test, about $0.034. A 4-frame 512×512 GIF cost about $0.017.

Can GPT Image 2.6 render text in other languages?

In our test both models rendered a Vietnamese headline with every accent correct, on four of four images. OpenAI does not publish a claim about non-Latin scripts, so test your language before relying on it.

Can I use GPT Image 2.6 in the Responses API?

No. The model pages list v1/images/generations and v1/images/edits as supported, and Responses, Chat Completions and Batch as not supported.