GPT Image 2.5 Flare vs Sunburst: Which to Pick Now That 2.6 Is Out

GPT Image 2.5 Flare vs Sunburst compared: speed, editing precision, cost and best uses. GPT Image 2.6 is now the latest and keeps the same Flare/Sunburst split.

Abstract split of swift cool-white and blue flare light on the left and a refined golden sunburst on the right, meeting at a bright center, with no text
Table of contents

When OpenAI released GPT Image 2.5 on September 8, 2026, it didn't ship one model. It shipped two, with names that sound more like energy drinks than software: Flare and Sunburst.

Update, November 2026: GPT Image 2.5 is no longer OpenAI's newest image model. On November 4, 2026, OpenAI released Images 2.6 with two new API models, GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst. The split works the same way: Flare for speed, Sunburst for detail and consistency across edits. This comparison covers the 2.5 versions, which you can still use on AI Studio. If you're starting fresh, try the GPT Image 2.6 generator, and see what 2.6 adds in our GPT Image 2.6 GIF examples.

The names don't tell you much, so here's the short answer: Flare is the fast, everyday model. Sunburst is the careful one for edits where details matter. For most people, Flare should be the default, and Sunburst is what you reach for when Flare keeps messing up the part of the image you asked it to leave alone.

That's the summary. The rest of this article covers where the differences actually show up, what each one costs, and a simple workflow that uses both.

What GPT Image 2.5 is

GPT Image 2.5 is the model behind ChatGPT Images 2.5. The ChatGPT release added a feature called Sketch, which lets you draw a rough layout and have the model turn it into a finished image.

For developers and image platforms, OpenAI offered the model as two API versions:

  • GPT-Image-2.5 Flare, which OpenAI called its fastest model at launch for high-quality, everyday image generation. OpenAI says it produces better images than GPT Image 2 with 50% lower latency.
  • GPT-Image-2.5 Sunburst, which OpenAI called its most capable model at launch for image generation and editing, built for workflows where editing precision matters most. The tradeoff, in OpenAI's words, is longer generation times.

Both take text and images as input, both support inpainting and the image edit endpoint, and both share the same quality settings, sizes, and rate limits.

Side by side

FlareSunburst
OpenAI's description at launchFastest model for high-quality everyday imagesMost capable model for generation and editing
Speed50% lower latency than GPT Image 2, per OpenAISlower; OpenAI says "longer generation times"
Where it's strongestVolume, drafts, social and product contentPrecise edits, retouching, combining several images
Quality settingslow, medium, high, xhigh, max, autoSame
Max sizeUp to 4K (3,840 pixels on the long edge)Same
OpenAI API token rates$8 per million image input tokens, $30 per million image output tokensSame
Credits on AI StudioAbout a quarter of Sunburst's costAbout four times Flare's cost

Image quality: where the gap actually shows

Both models are good. The question is where Sunburst's extra care is worth waiting for.

DataCamp's launch write-up cited leaderboard scores that put Sunburst ahead of Flare in both text-to-image and image editing, with a wider gap in editing (1,520 vs. 1,491) than in text-to-image (1,421 vs. 1,399). DataCamp didn't name the leaderboard, so treat those numbers as a rough signal, not a verdict. The pattern matches how OpenAI positions the two models, though: the difference is mostly about editing, not first drafts.

In practice, you'll notice Sunburst most when you ask for changes to an existing image, like:

  • "Change the jacket to navy, keep everything else exactly the same."
  • "Put this product into this kitchen scene and match the lighting."
  • "Combine these four photos into one group portrait."

These are the jobs where a model either respects what you asked it to keep or quietly redraws it. A shifted logo, a different face, or a label in the wrong place is the gap between a usable asset and a do-over.

For a first draft, like a blog header, a social post, or a mood board, Flare usually gets you there faster, and the difference is hard to spot at phone size.

Neither model solves everything. In DataCamp's hands-on test, dense small text and typography still hit the usual limits, and edits that moved several objects at once could leave artifacts. If your design is mostly text, plan to set the type yourself.

Speed: why it matters more than it sounds

OpenAI's 50% latency claim for Flare comes from its own testing, so your results will vary with size and quality settings. The principle holds, though: fast models change how you work.

When an image comes back quickly, you try more ideas. You test five compositions instead of one and keep the best. When each image takes noticeably longer, you start writing longer prompts and hoping for the best.

That's why Flare is the better exploring model, even if Sunburst might win any single head-to-head comparison.

Price: same rates, different bills

This is where things get confusing, so let's take it slowly.

On OpenAI's API, Flare and Sunburst have identical token prices: $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens, with batch processing at half price. Those are the same rates as GPT Image 2, and OpenAI lists the same token rates for GPT Image 2.6 Flare and Sunburst.

What changes your bill most is the quality setting. According to OpenAI's calculator, as reported by Kanaries, a 1024×1024 image produces these output costs:

QualityApprox. output cost
low$0.006
medium$0.013
high$0.053
xhigh$0.094
max$0.211

Those figures cover output tokens only. Your prompt and any reference images add input cost on top. So on the API, "Sunburst at medium" can cost less than "Flare at max."

On AI Studio, pricing works differently. You pay in credits, and the GPT Image 2.5 page says Flare uses about a quarter of the credits Sunburst does for the same settings. Higher resolutions cost more on both. The Generate button shows the exact cost before you run anything, and you can compare plans on the AI Studio pricing page.

The practical takeaway is the same either way: don't use the expensive setting for drafts.

Which one for which job

Choose Flare for:

  • Social posts, thumbnails, and blog images
  • Exploring several visual directions quickly
  • Batch work, such as 40 product angles or a full set of illustrations
  • Quick prototypes and mockups
  • Anything where "very good, right now" beats "perfect, later"

Choose Sunburst for:

  • Retouching a hero image you already like
  • Product photos where the label, shape, and color must stay exact
  • Combining several reference images into one scene
  • Final campaign assets that will be printed or shown large
  • Any edit where Flare keeps changing things you asked it to keep

A two-step workflow that uses both

  1. Explore with Flare at low or medium quality. Generate 6 to 10 variations, and change one thing at a time between rounds so you learn what actually helps.
  2. Pick the winner. Choose the composition, not the details. Details are fixable.
  3. Refine with Sunburst. Feed the winning image back in as a reference and make specific edits: "Keep the composition and the model's pose. Change the background to a concrete studio wall. Keep the shadow under the bottle."
  4. Render the final at high resolution. Only go to 4K once you're done changing things. Resolutions above 2560×1440 are marked experimental in OpenAI's docs, so check the details at full size.

This is how a good photographer works too: lots of quick test shots, then careful work on the one that matters.

Prompt tips that help with both

  • Say what to keep, not just what to change. "Keep the face, the pose, and the logo position" does more than "edit the background."
  • Describe light like a photographer. "Soft window light from the left, late afternoon" beats "nice lighting."
  • Give text its own line. Put any words you need in quotation marks, keep them short, and check the spelling in the output.
  • Number your references. When combining images, say which one supplies what: "Use image 1 for the person, image 2 for the jacket, image 3 for the location."
  • Match the aspect ratio to where it'll be used. A 9:16 story image and a 16:9 header are different compositions, not crops of each other.

How to try both

You can use both versions in the GPT Image 2.5 image generator on AI Studio. Choose Flare or Sunburst, choose 1K, 2K, or 4K, and pick from 15 aspect ratios, from 1:1 and 16:9 to tall 1:3 and wide 21:9. Image-to-image mode accepts up to 16 reference images, which is useful for the multi-image work where Sunburst earns its keep.

If you want to see how GPT Image 2.5 compares with other models, AI Studio's AI image generator also includes options like Nano Banana, Seedream, and Z Image. It's worth running one of your real prompts through two or three of them. Models have different instincts about style, and the "best" one is often just the one whose taste matches yours.

The bottom line

Start with Flare. It's fast, cheap at low quality, and good enough for most images most people need. Switch to Sunburst when you're editing something that already works and you can't afford to have the model "improve" the parts you love.

The same rule carries over to GPT Image 2.6, which is now OpenAI's current version: Flare for fast drafts, Sunburst for careful edits.

Sources

Create with AI Video Studio

Turn your idea into a video

Explore the video generator and start creating.

Generate a video

Written by

AI Studio Team