beckon
Beckon blog

GPT Image 2.6 Flare vs Sunburst vs 2.5: Blind Test Results

ยท Beau Johnson

In shortIn my blind test, quality across GPT Image 2.6 Flare, 2.6 Sunburst and 2.5 Sunburst came out close to a three-way tie. Flare was about 4.5x faster than 2.5, and 2.6 Sunburst was the only model to get all five chained edits right. I'd switch: Flare for volume, Sunburst for edits.

OpenAI released Images 2.6 as two API models, Flare and Sunburst. On Arena's text to image leaderboard, though, the older GPT Image 2.5 Sunburst still sits at number one. In this video I send the same prompts to 2.6 Flare, 2.6 Sunburst and 2.5 Sunburst at the same moment, have Gemini and Claude judge the results blind, and tell you which one I'd put in an ad pipeline. It's for builders who generate images through an API and want to know whether switching is worth it.

What I cover in this video

  • 0:00 One ad prompt, three models, one stopwatch
  • 0:47 What's new in 2.6: Flare, Sunburst, GIFs, SVG, markup and complexity
  • 3:05 The Arena scoreboard and why the old model is still on top
  • 3:52 How I set up the blind test
  • 5:24 Five edits in a row on a product shot
  • 7:39 Low complexity and the speed round
  • 8:26 The business round: ads, a GIF, an SVG logo and a markup edit
  • 9:56 The verdict

What's new in Images 2.6

This section is what OpenAI says, not what I tested. Images 2.6 shipped on November 4 in ChatGPT and Codex. In the API it's two models:

  • GPT Image 2.6 Flare: built for fast iteration and clean, professional layouts like posters and product shots.
  • GPT Image 2.6 Sunburst: built for detail and for staying consistent when you edit the same image again and again.

OpenAI says it generates images 50% faster than before, and that Flare can produce an image in as little as 6 seconds at low quality. The new features are animated GIFs with adjustable playback speed, saved reference images in ChatGPT, markup edits in the API (you draw on the image and it changes what you marked), SVG export, and a new complexity setting.

Complexity controls how much planning the model does before it makes the image. Low is for quick edits, and medium or high is for busy scenes. Easy to miss: if you don't set it, both models default to high, the slow setting, even though OpenAI recommends low for most edits.

The scoreboard: why 2.5 is still number one

Arena runs blind matchups. Two anonymous models answer a prompt, you pick the better one, and the votes become a rating, a bit like a chess rating. (If you want more background on how leaderboards like this work, see Coding Agent Benchmarks Explained.) At launch, the text to image board looked like this:

  1. GPT Image 2.5 Sunburst: 1,427
  2. GPT Image 2.6 Sunburst (Medium): 1,409
  3. GPT Image 2.5 Flare
  4. GPT Image 2.6 Flare (Medium)

Each old model sits above its own new version. On the image edit board as of November 8, 2.5 Sunburst leads by 43 points. Arena's speed numbers go the other way. Median time per image dropped from 36.9 s to 19.4 s for Sunburst and from 21.7 s to 13.4 s for Flare.

Some context: 18 points is a small gap, all four rows are still marked preliminary, and 2.6 Sunburst has about 2,700 votes against more than 58,000 for 2.5 Sunburst. It's early. So I ran my own test.

How I tested, and what happened in each round

All three models ran through the same third-party provider, which had 2.6 live at the time. Every prompt went to all three models at the same moment, so none of them got a quieter server. I used the same resolution for all three and ran 2.6 at medium quality, the setting Arena timed. 2.5 ran at the provider's default because the provider doesn't let you change its quality. Each image was stripped to an unlabeled file, shuffled, and scored by two judges that aren't OpenAI models, Gemini and Claude, using the same scoring sheet every time.

Round one: text

The prompts were a 14-line chalkboard cafe menu, an event poster, the small print on the back of a coffee bag, and a four-step infographic. All three models spelled every word correctly. The only misses were punctuation: 2.5 added an extra period on the coffee label, and 2.6 Sunburst dropped the periods after the infographic step numbers. 2.5 won the round by a hair, taking the menu and the infographic.

Round two: five edits in a row

This is the round I care most about for a real business. Each model edited its own mug product shot five times: put it on a cafe table, change the color to navy, add a hand holding it, hang a gift tag that says "For Sam", then make it evening. Each edit produces a whole new picture from the previous one, so small drift adds up over the chain.

Both judges said 2.6 Sunburst was the only model to get all five edits right. Flare and 2.5 both kept morning sun patches on the table for the evening edit, and Flare never added a lamp. Sunburst added a lamp and a dusk window. It was also quicker: about 60 s per edit for 2.5, about 40 s for Sunburst, and 19 s for Flare, all with complexity left on the default high.

Round three: a busy scene

A night market with 11 required elements, including three food stalls with signs, a red bike, a golden retriever, a woman in a yellow raincoat with a clear umbrella, and a clock at 9:15. Both 2.6 models got all 11, and 2.5 averaged 10.5. 2.5 and Flare both put a real watch brand name on the clock that I never asked for. That's the kind of thing you need to catch before an ad goes out.

With complexity on low, both 2.6 models still got 11 out of 11. Sunburst went from 44.9 s to 18.9 s, and Flare actually got slower, going from 12 s to 19 s. That's one run each, so treat it with caution. Still, if you use Sunburst and your scene isn't very crowded, low complexity looks like free speed.

Round four: speed

Across 12 text to image runs per model, the medians were 55.0 s for 2.5 Sunburst, 27.6 s for 2.6 Sunburst and 12.1 s for Flare. That makes Flare about 4.5x faster than 2.5. The roughly 2x gain for Sunburst matches Arena. I never saw the 6-second figure, though. Flare at low quality came in at 12 s on all three tries through my provider.

The business round

For a made-up brand, a terracotta travel mug, I made a square feed ad, a vertical story ad and a wide banner. All three models got every word right, the judges split their picks between Flare and Sunburst, and 2.5 wasn't picked once. Then I tried the new features: a 12-frame looping GIF of the mug with steam, an SVG logo, and a markup edit where I circled the lid in red and wrote "gold". Both 2.6 models changed only the lid, and Flare did it in 14 seconds. I only ran markup on 2.6.

Where it still breaks: the SVG logos were basic and both drew a handle my mug doesn't have (in fairness, my prompt never said "no handle"). There was also the unrequested brand name, and Sunburst threw one server overloaded error on the first GIF during launch week, which worked on retry. Treat the output as early drafts and check everything before it ships.

Key takeaways

  • Quality is roughly a three-way tie. Out of 16 best-image picks, Flare got 6, 2.6 Sunburst got 6 and 2.5 Sunburst got 4.
  • Use Flare for volume. It was about 4.5x faster than 2.5 Sunburst and held its own on quality.
  • Use Sunburst for repeated edits. It was the only model to get all five chained edits right.
  • Set complexity yourself. It defaults to high, and low cut Sunburst's busy-scene time from 44.9 s to 18.9 s in one run.
  • I'd move off 2.5 Sunburst. It was the slowest, can't do GIFs or SVGs, and didn't win the edits in my test, even though it still leads on Arena.

FAQ

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

OpenAI positions Flare for fast iteration and clean layouts like posters and product shots, and Sunburst for detail and consistency across repeated edits. In my test, Flare was the fastest and Sunburst was the only model to get all five chained edits right.

Why is GPT Image 2.5 Sunburst still ranked above 2.6 on Arena?

At launch, 2.5 Sunburst led 2.6 Sunburst by 18 points on text to image, which is a small gap. All four rows were still marked preliminary, and 2.6 Sunburst had about 2,700 votes compared with more than 58,000 for 2.5.

What does the complexity setting do in GPT Image 2.6?

It controls how much planning the model does before generating. Both models default to high, even though OpenAI recommends low for most edits. In one run, low complexity cut Sunburst's busy-scene time from 44.9 s to 18.9 s and kept all 11 elements.

How fast is GPT Image 2.6 Flare?

Across 12 runs in my test, Flare's median was 12.1 seconds per image, about 4.5x faster than 2.5 Sunburst's 55.0 seconds. I never saw OpenAI's 6-second figure through my provider. Flare at low quality took 12 seconds on all three tries.

Run your agents in Beckon

A workspace for your coding agents. Get more work done in less time.

Download BeckonSubscribe on YouTube
Full video transcript

0:00Same ad, same prompt. Three image models fired at the exact same second. GPT image 2.6 Flare finished in 12 seconds. 2.6 Sunburst took 27 1/2. And the old one, 2.5 Sunburst, almost 55 seconds. So, the new Flare is about 4 1/2 times faster than the old one. Easy win, right? Well, here's the twist. On Arena's textto-image leaderboard, the old 2.5 Sunburst is still sitting at number one above the brand new 2.6. So, which one should you actually be using? So, in this video, we're going to be breaking down what's actually new in 2.6, what the independent scoreboard says, and then our own head-to-head text, product edits, busy scenes, speed,

0:47and a full ad set for a brand we made up. and I'll tell you which one I'd put in my ad pipeline and if it's even worth switching. Let's jump right into it. So, OpenAI shipped ChatGpt images 2.6 on November 4th and it's in ChatGpt and Codeex on every plan. Everything in this section is what OpenAI says, not what we tested. We'll get to that. In the API, 2.6 is actually two models. Flare is for fast iteration and clean professional layouts like posters and product shots. Sunburst is for detail and for staying consistent when you edit the same image over and over. Open AAI says it generates images 50% faster than before and that Flare can do an image in as little as 6 seconds at low quality. And

1:34then here's the new stuff. It can make animated GIFs and you can change the playback speed. In chat GPT, you can save reference images, so yourself, your dog, a character, and reuse them without uploading again. Markup edits came to the API. So you draw on the image and it changes what you marked. It can export SVG, so actual vector files. And there's a new setting called complexity. So what does complexity actually do? OpenAI says it controls how much planning the model does before it makes the image. Low is for quick edits. Medium or high is for busy scenes with a lot of things in them. And here's something most people don't think about. If you don't set it, both models default to high, the slow setting, even though OpenAI recommends

2:20low for most edits. So, think of them as two cars from the same team with a different tune. Flare is tuned for a clean result fast. Sunburst is tuned to hold on to the details on your fifth edit. Every night, Night Crew researches what your market is talking about, writes the script, fact checks it, voices it in your voice, and animates it. In the morning, you approve it and it posts to YouTube, Instagram, Tik Tok, and Facebook. Nightcrew. Nightcrew.studio. So, what does the independent scoreboard say? Arena runs blind matchups. You type a prompt, two anonymous models answer, you pick the better one, and only then do you see who made what. All those votes turn into a rating. Kind of like a chess rating. Beat a model that's rated higher than you, and you gain more

3:05points. And here's where it gets weird. At launch, Arena's text to image board had 2.5 sunburst at number one with 1427. 2.6 sunburst at medium, number two, 1409. Number three was the old 2.5 flare, and the new flare at medium was number four. So each old model sits above its own new version. And on Arena's image editing board as of November 8th, it's the same story except there 2.5 Sunburst leads by 43 points. But look at speed. Arena measured the median time per image. 2.5 sunburst took 36.9 seconds. 2.6 sunburst at medium took 19.4. On Flare, it went from 21.7 down to 13.4. So, the new ones are a lot

3:52faster, and they're rated a little lower. Now, here's what this is all about. On text to image, 18 points is a small gap, and all four of these rows are still marked preliminary. 2.6 6 sunburst has about 2700 votes. 2.5 Sunburst has over 58,000. So, it's early. Some people on X are saying 2.6 is worse than 2.5, but that's opinions, not a test. So, we ran our own. So, here's how we set it up. We ran all three models through the same provider, Key.AI, which had 2.6 Live. When we tested it, every prompt went to all three models at the exact same moment. So, nobody got a quieter server. Same resolution and 2.6 at medium quality, the same setting arena timed.

4:382.5 ran at whatever default key serves it at because keyi doesn't let you change its quality. Then we scored it blind. Every image got stripped down to a plain file with no labels, shuffled, and handed to two judges that are not OpenAI models, Gemini and Claude, with the same scoring sheet every time. They never knew which model made what. Four rounds. text, edits, a busy scene, and speed. Then the business round. Let's see what we got. Round one is text. A chalkboard cafe menu with 14 lines, an event poster, the tiny print on the back of a coffee bag, and a four-step infographic. And honestly, text is basically solved, at least on our prompts. All three spelled every single

5:24word right. Every word. The only misses were punctuation. 2.5 put an extra period on the coffee label and 2.6 Sunburst dropped the periods after the step numbers on the infographic. 2.5 took the menu and the infographic. Flare took the label. Sunburst took the poster. So 2.5 wins round one, but by a hair. Round two is the one I care about most for a real business. We made a product shot of our mug and then each model edited its own picture five times in a row. Put it on a cafe table. Change the color to navy. Add a hand holding it. Hang a gift tag that says for Sam. Then make it evening. Here's how drift actually works. Every edit, the model makes a whole new picture from the last

6:09one. So, a tiny change in edit one gets carried into edit two and three. And by edit five, your product can look like a cousin of itself. That's what we're scoring. Does the same mug survive? and Sunburst was the only one that did all five edits right according to both judges. Flare and 2.5 both slipped on the last one. You asked for evening and they both kept the morning sun patches on the table and Flare never put in a lamp. Sunburst put in a lamp and a dusk window. So, OpenAI's pitch for Sunburst consistency across edits held up in our test. Both judges picked Sunburst's chain as the best and it did it faster. 2.5 took about 60 seconds an edit. Sunburst took about 40 and flare took

6:5419. And that's with complexity on the default high. So, Sunburst could go even faster on edits. Round three is the busy scene. A night market with 11 things that all have to be there. Three food stalls with signs, a red bike on a lamp post, a golden retriever, string lights, a woman in a yellow raincoat with a clear umbrella, an open late chalkboard, a clock at 9:15, and puddles. Both 2.6 models got all 11. 2.5 got 10 1/2 on average, so pretty close. One thing that did slip, though, 2.5 and Flare both put a real watch brand name on the clock that we never asked for. That's the kind of thing you have to catch before an ad goes out. Then we flipped complexity to

7:39low. Still 11 out of 11 on both. Sunburst went from 45 seconds down to 19. Flare actually got slower from 12 to 19. That's one run each. So take it with a grain of salt. But if you're on sunburst and your scene isn't crazy, low complexity looks like free speed. Round four. Speed. Across 12 textto image runs per model. The median for 2.5 sunburst was 55 seconds. 2.6 sunburst was about 28 and flare was 12 seconds. That's about 4 1/2 times faster than 2.5 which is insane. Arena said about two times faster for Sunburst and we got two times faster for Sunburst. So that lines up. What we didn't see was 6 seconds. We ran

8:26Flare at low quality three times and got 12 seconds every single time. So that 6-second number is open AIS on their own API. And through our provider, we never saw it. So here's how I'd actually use this. Our madeup brand is Pillow, a terracotta travel mug. One brand kit, three colors. The ad set, a square feed ad, a vertical story ad, and a wide banner for the website. All three models nailed every word on all three ads, and the judges split their picks between Flare and Sunburst. 2.5 didn't get picked once on the ads. Then the new stuff, a looping GIF of the mug with steam coming off it. From what our provider's docs say, the model draws all the frames at once as one grid, like a

9:11sheet of stickers, and then it gets cut up and played in order. 12 frames and it loops. Then an SVG logo. The API makes a normal picture first and then a text model writes the vector code from it. So you get a real file you can open and edit. and the markup edit. We circled the lid in red, wrote gold, and both models changed the lid and nothing else. Flare did it in 14 seconds. Open AAI says markup works on 2.52, but we only ran it on 2.6. That's an afternoon of a designer's time, done in a few minutes. And here's where it still breaks. The SVG logos are pretty basic, and both drew a mug with a handle when ours doesn't have one. To be fair, our logo prompt never said no handle. The evening

9:56edit, we talked about a real brand name showing up where we didn't ask for one. And on launch week, Sunburst threw one server overloaded error on our first GIF. The retry worked. So, treat it as the first 10 versions, not the last one. Check everything before it goes out. So, the final score on quality, it's basically a three-way tie. Out of 16 picks for the best image, the judges picked Flare six times, Sunburst six times, and 2.5 four times. That lines up with Arena, where the textto image gap is small and the votes are still coming in. So, which one should you use? If you're pumping out ad variations, thumbnails, anything where you want a lot of options fast, use Flare. Four and a half times faster than 2.5 Sunburst in

10:43our test, and it held its own on quality. If you're editing the same product over and over, use Sunburst. It was the only one that got all five edits right. And 2.5 Sunburst, honestly, I can't find a reason to stay on it for this kind of work. It was the slowest by a mile, and it can't do the GIFs or the SVGs. It still leads on arena by a little on text to image and by more on editing, but in our test, it didn't win the edits. I'd switch. So, which one are you putting in your ad pipeline? flare or sunburst, or are you sticking with 2.5? Tell me down below. And if this saved you a weekend of testing, a thumbs up really helps the channel. Have a blessed day.