GPT Image 2.6 Tested: Sunburst, Flare and the New Dial

Every image model sounds great in its launch post. The real test is boring and specific: does it put the label on the right mug, does it leave the cat alone when you ask for a new glaze, and does it come back fast enough that you'll actually iterate? We ran that test the morning GPT Image 2.6 showed up on each::labs.

OpenAI released GPT Image 2.6 on November 4 as two models, Sunburst and Flare, and both are now live on each::labs for text-to-image and editing. This is not a feature tour of the announcement. It's what we saw when we ran the same brief through both models, at two complexity settings, next to the GPT Image 2.5 model many of you already use.

What's Live on each::labs Right Now

Four pages, one family. GPT Image 2.6 Sunburst Text to Image and Sunburst Edit are the detailed pair. Flare Text to Image and Flare Edit are the fast pair. OpenAI's own image guide puts the split in one line: "Sunburst for detailed visual work and Flare for fast iteration."

The request schema on each::labs is the useful part to know before you write a line of code. Every GPT Image 2.6 page takes a prompt of up to 32,000 characters, a new complexity field (low, medium or high), a quality field that runs from low through medium and high up to xhigh and max, as many as 10 images in a single request, and transparent backgrounds when you ask for PNG or WebP. Sizes go from 1024x1024 up to 3840x2160 and 2160x3840, with anything above 2560x1440 marked experimental. The edit pages accept up to 16 source images.

One thing to set expectations on. OpenAI's API also added markup-guided edits, SVG output and animated GIF output with this release. Those aren't inputs on the each::labs pages today: you edit with a prompt and source images, and you get back PNG, JPEG or WebP. If you've read about drawing red circles on a photo, that workflow isn't part of what we tested here.

The Test: One Brief, Five Runs

We wanted a prompt that punishes sloppy reading. So we asked for a cluttered ceramic studio shelf in morning light, twelve hand-thrown mugs in different glazes, a handwritten paper label on the third mug reading "BATCH 07", a sleeping grey cat between the fifth and sixth mug, and dust in the light beam. Editorial product photo, 35mm, shallow depth of field.

A cluttered ceramic studio shelf, morning window light: twelve hand-thrown mugs in different glazes, a handwritten paper label on the third mug reading 'BATCH 07', a sleeping grey cat between the fifth and sixth mug, dust in the light beam. Editorial product photo, 35mm, shallow depth of field.

That's three spatial instructions (third mug, fifth and sixth mug, a count of twelve) plus a short line of text that has to be spelled right. We ran it on Flare and Sunburst at low and high complexity, all at medium quality and 1536x1024, and once more on GPT Image 2.5 Flare for a baseline.

Every run spelled "BATCH 07" correctly and put it on the third mug. That's the headline nobody puts in a press release, and it's the one that saves you a re-roll. No model gave us exactly twelve mugs, which tells you something about counting in a cluttered scene: treat exact numbers as a hope, not a spec.

Sunburst Reads the Whole Sentence

The cat was the real exam. Both Sunburst runs laid out a single long shelf of mugs and placed the cat right after the fifth one, exactly where the prompt asked. Flare got the label right but treated the cat more loosely: at low complexity it curled up after the fourth mug, and at high complexity it rebuilt the scene as two stacked rows, which made "between the fifth and sixth" ambiguous.

That matches OpenAI's description almost word for word. Sunburst is the one for detailed visual work, the shots where several instructions have to land at once. If your brief has positions, counts and text in the same sentence, start on GPT Image 2.6 Sunburst and save yourself the argument.

Look closer and the two Sunburst images differ in how they spend their effort. The high-complexity version added more of the studio around the shelf: a sketch of a mug clipped to a board, a second shelf of bowls, a fuller light beam. The low-complexity version was tighter and plainer. Neither invented stray text. Our GPT Image 2.5 baseline, by contrast, also got the label and the cat right, but it filled the wall with extra writing we never asked for, from a poster to book spines. Small thing. It's exactly the kind of small thing that gets an image bounced in review.

Turn the Complexity Dial With a Reason

Complexity is the genuinely new control in GPT Image 2.6, and it's easy to misread. It isn't quality. OpenAI's guide describes it as how much the model plans before rendering, separate from the quality setting that controls how finely the image is rendered. If you leave it out, both models default to high.

OpenAI is unusually direct about when to use which. Lower complexity "can suit focused edits, such as changing a background or a sign." Higher complexity "can help with dense text, complex layouts, and relationships between objects." And the warning matters as much as the advice: higher complexity "can increase latency and introduce unwanted changes, so compare results for your task."

Our runs backed that up in miniature. High complexity on Sunburst gave a richer, more composed studio. It also took more liberties with the background. For a hero image, that's a gift. For a product shot where the art director has already approved the set, it's a risk. Our take: run low or medium when the job is to change one thing, and high when the job is to arrange many things.

Edits That Leave the Cat Alone

Then we edited. We took the Flare image and sent it to both edit models with one instruction: change only the glaze of the third mug from the left to deep cobalt blue, and keep everything else identical, including the label text, the cat, the light and the framing. Medium complexity, medium quality.

Change only the glaze of the third mug from the left to deep cobalt blue. Keep everything else identical: the label text, the cat, the light, the framing.

Both edits preserved the scene. Same shelf, same cat mid-nap, same plant, same dust in the light, same label still reading "BATCH 07". The difference was interpretation. That Flare image had two rows of mugs, and GPT Image 2.6 Flare Edit recolored the third mug of the top row, while GPT Image 2.6 Sunburst Edit recolored the third mug of the bottom row, the one wearing the label. Both are defensible readings. Neither is the one you'll get every time.

The lesson is old but worth repeating. An edit is only as precise as the way you point. "Third mug from the left" was clear when we wrote it and ambiguous once the shelf had two rows. Point with something unique instead: "the mug with the BATCH 07 label", "the mug next to the cat". Anchoring an edit to a visible, one-of-a-kind detail is the easiest precision upgrade there is.

How Fast GPT Image 2.6 Feels in Practice

Speed claims are the easiest thing to inflate, so here's only what we measured. A handful of calls can't give you a benchmark, and we won't pretend otherwise. What we can tell you is what our runs felt like end to end on each::labs, queue included: every GPT Image 2.6 generation and edit came back in somewhere between about 10 and 30 seconds at medium quality, and the fastest result of the morning was a Flare run.

Single runs are noisy, and our low-complexity Flare call was actually slower than the high-complexity one, which says more about the queue that minute than about the model. If speed decides your pipeline, time it on your own prompts over a few dozen calls. The pattern that held up across our runs is simple: Flare for the loop where you react and adjust, Sunburst when the result is the one you'll ship.

A Workflow That Uses Both Models

The cleanest way to work with GPT Image 2.6 on each::labs is to stop choosing one model. Explore on Flare with low or medium complexity and several images in each request (the pages allow up to 10) to find the composition. Move the winning prompt to Sunburst at high complexity for the final. Then fix anything small with an edit at low complexity, pointing at a unique detail, so the rest of the frame stays put.

Because both models sit behind the same each::labs API, switching is a model id change, not a new integration. If you're already on GPT Image 2.5, the same is true for moving up: the fields you know still work, and complexity is one optional field on top. Chain the final image into an image-to-video model or a full each::labs flow and the still becomes the first frame of a campaign.

Prompt structure carries over too. Everything in our GPT Image 2.5 prompting guide applies, and if you want to see where the previous generation already earns its keep, the GPT Image 2.5 use cases post is the map. For how GPT Image stacks up against other editors, our roundup of the best AI image editing models is the place to compare.

Who Should Switch to GPT Image 2.6 Today

If your images carry text, positions or several objects that relate to each other, switch now and start on Sunburst. Packaging mockups, shelf and retail scenes, menu boards and infographics are where reading the whole sentence pays off. Teams that generate in volume, social and performance marketers testing many variants, should move their exploration loop to Flare and treat complexity as a new knob to test, not a setting to max out.

If you mainly retouch existing photos, the edit pages are the draw. Both preserved our scene through a targeted change, and room for up to 16 source images in one request gives you space for references. And if you're happy with GPT Image 2.5 output for simple, single-subject images, there's no rush. Run your top ten prompts through GPT Image 2.6 side by side and let the results make the call.

Frequently Asked Questions

Is GPT Image 2.6 available on each::labs?

Yes, since this week. Both models are live, each with a text-to-image page and an edit page: Sunburst Text to Image, Sunburst Edit, Flare Text to Image and Flare Edit. You call them through the same each::labs API as every other model in the catalog.

What does the complexity setting in GPT Image 2.6 do?

It controls how much the model plans before it renders, which is different from quality. Low suits focused edits such as swapping a background or a sign, high suits dense text, complex layouts and scenes with many related objects. It defaults to high, and OpenAI notes that higher settings can add latency and unwanted changes, so test both on your own prompts.

Should I use GPT Image 2.6 Sunburst or Flare?

Think in phases rather than winners. Flare is the fast iteration model, so it belongs in the exploring part of the job. Sunburst is built for detailed work, and in our test it followed every spatial instruction in a crowded prompt, so it's the one for final images and briefs with positions, counts and text.

Can GPT Image 2.6 make SVGs and GIFs on each::labs?

Not today. OpenAI added SVG output, animated GIF output and markup-guided edits to its own API with this release, but the each::labs pages currently return PNG, JPEG or WebP and take a prompt plus source images. Raster generation, editing, complexity and transparent backgrounds are all there now.