Krea 3 Is Out of Beta: 8 Things to Know About Ultra

Fast image editors have always come with an unspoken deal. You get the result in seconds, and you accept that the hardest edits come back a little wrong. Five products in one scene, a character who has to stay the same person, a label that can't move a pixel: that's where speed used to run out.

On October 20, Krea changed the terms. Krea 3 officially left beta, Krea published a full set of leaderboard results, and its slowest, most careful mode, Ultra, arrived in the API. Ultra is already a setting on the each::labs Krea 3 endpoint. If you've been using Krea 3 for quick edits, here are eight things worth knowing now that it's out of beta, starting with what Ultra actually does differently.

1. Krea 3 Is Out of Beta, and That's More Than a Label

Krea introduced Krea 3 on October 15 and moved fast after that. Its release notes show API early access for beta workspaces on the 15th, Krea 3 as the default image model in the Krea app on the 16th, every Krea workspace able to call the API on the 19th, and "Krea 3 is out of beta" on the 20th, together with Ultra in the API and Krea 3 showing up in Krea's ChatGPT and iOS model pickers.

Why care about a label? Because "beta" is a reason teams hold back. Out of beta means the model has a stable place in Krea's lineup and the API modes are settled. Krea also renamed its old thinking modes in the process (the earlier names stop working on October 23), so if you saw older Krea 3 examples using different mode names, they now map to fast and high.

2. Ultra Doesn't Think Harder. It Tries Several Times and Picks

This is the bit most people get wrong. Krea's developer docs describe three modes. Fast "quickly expands the prompt, then generates." High uses "a stronger reasoning engine to plan more complex edits from the prompt and the reference images, then generates." And Ultra "generates several candidates, reviews them against your prompt and references, and returns the best one."

So Ultra isn't a bigger brain. It's a built-in art director: several attempts, a review pass against your brief, one winner. That matters because it changes when Ultra is worth it. If your edit is simple, fast already gets it right and Ultra adds only waiting. If your edit has many ways to go subtly wrong, having the model reject its own near-misses is exactly what you want.

3. The Benchmarks Put Krea 3 Ultra in the Top Five

Krea's benchmark post collects four public leaderboards, read on October 19 and 20. As Krea reports them, Krea 3 Ultra sits fourth on Arena's multi-image edit board (1435, preliminary, from 2,991 votes) and eighth on Arena's single-image edit board. On Artificial Analysis Image Editing it ranks fifth with 1151, and Fast mode ranks eleventh with 1125. On Design Arena's image editing board, the Ultra Thinking variant is fourth of 51.

Two honest caveats come from Krea itself. The boards use different scales, so the numbers shouldn't be compared across them, and the Arena scores were still marked preliminary. What the spread does show is consistent: Krea 3 is a top-tier editor, and Ultra is what gets it there.

4. Multi-Image Composition Is Where Ultra Pulls Ahead

Look at where Krea 3 ranks best. Its strongest placement is on multi-image editing, and in Artificial Analysis's category breakdown Ultra places second in Animation and Gaming, third in Composition and Framing, and fourth in Reasoning-Based Edit. Those are the jobs where you're asking the model to understand how several things fit together, not just to change one.

The each::labs endpoint takes up to eight images in image_urls. One image means edit it; two or more means combine them, for example a product and a scene. Refer to them by position in the prompt so the model knows which is which:

Place the sneaker from the first image on the wet city pavement from the second image at night, lit by the neon sign from the third image. Keep the sneaker's logo, laces and colours exactly as in the first image. Low camera angle, shallow depth of field.

That's an Ultra brief. Several references, one product that must not change, and a lighting cue borrowed from elsewhere.

5. Fast Still Wins More Often Than You'd Think

Here's the counterintuitive part of Krea's own data. In the Artificial Analysis categories, Fast mode beats Ultra on Live-Action Film and on Scene and Style Edit. Ultra places eighteenth on Scene and Style Edit. More planning doesn't help every kind of edit, and for broad restyles it can actively get in the way.

Krea also says plainly where the whole model is still weaker: small and long text, full style changes, and perfect identity across many edits. So the right default isn't "always Ultra." It's fast for most edits and exploration, high for complex edits with many changes, and Ultra for the few edits that carry the most weight. The each::labs field description puts timing on it: fast lands in about 3 to 8 seconds, high in about 10 to 30.

6. Ultra Has Rules: Bring a Reference, Render at 2K, Plan for Minutes

Ultra is an editing mode. Krea's release notes say it needs at least one reference image, and its docs note that a text-only Ultra request simply runs as fast. On each::labs, Ultra only works with resolution set to 2k, which matches Krea's statement that Ultra output is always 2K.

It's also slow by design. Krea warns that an Ultra job can take several minutes and advises using a webhook or a long polling timeout. And Krea documents that Ultra skips some features available in the other modes, including regional edits and negative prompts. Put together: use Ultra in pipelines where a final asset matters more than turnaround, and don't put it behind a button a user is staring at.

{
  "prompt": "Place the perfume bottle from the first image on the stone ledge from the second image at golden hour. Keep the bottle, label and cap exactly as they are.",
  "image_urls": ["https://your-bucket.example/bottle.png", "https://your-bucket.example/ledge.jpg"],
  "thinking_mode": "ultra",
  "resolution": "2k",
  "aspect_ratio": "4:5"
}

7. Krea 3 Against GPT Image 2 and Nano Banana Pro

Krea published two head-to-head posts, and they're more balanced than most vendor comparisons. Against Nano Banana Pro, Krea reports a lead on all three boards it cites, while conceding that Google's model is ahead on 4K output (Krea 3 tops out at 2K), Search grounding and long text. Against GPT Image 2, Krea 3 Ultra edges ahead on Artificial Analysis, while GPT Image 2 leads on Arena's multi-image board and Design Arena, and Krea itself recommends GPT Image 2 for dense or multilingual text.

That's the useful reading: pick by the job. Product and character consistency in a new scene, at speed, is Krea 3's ground. Heavy typography goes to GPT Image 2. 4K finals and long text go to Nano Banana Pro. All three are on each::labs, so you can run the same brief through each and keep the winner. Our guide to the best AI image editing models in 2026 goes wider than these three.

8. Style References and LoRAs Are Still on the Way

One thing hasn't landed yet. Krea's comparison posts both say dedicated style references and LoRA training are "coming soon," and its release notes show style reference and LoRA requests on Krea 3 returning a coming-soon response. If your workflow depends on locking a house style through a trained LoRA, that piece isn't part of the out-of-beta release.

Until then, carry style the way the model already handles well: through the reference images themselves and a precise prompt. Our Krea 3 prompting guide covers how to write those instructions, and Krea 3 use cases shows how teams already structure their revision loops around it.

How to Use Krea 3 Ultra on each::labs

The Krea 3 Image Edit page exposes everything above: thinking_mode (fast, high or ultra), model (berry for the highest quality, jam for speed with quality close to berry), up to eight image_urls, 1K or 2K output, ten aspect ratios and a seed for repeatable results. Leave the images out and it works as text-to-image.

A workflow that respects how Ultra behaves looks like this: explore directions on jam with fast, lock the layout with high, then send only the final, reference-heavy composites to Ultra. Wire it as an each::labs flow with a webhook at the Ultra step, and the minutes-long render happens in the background while your team moves on. Everything else in the image editing catalog sits behind the same API key if a step needs a different model.

Frequently Asked Questions

What does Krea 3 Ultra mode do?

It generates several candidates, reviews them against your prompt and references, and returns the best one at 2K, per Krea's docs. That review step is why it's slower, and why it helps most on complex, multi-reference edits rather than simple ones.

Is Krea 3 out of beta?

Yes. Krea's release notes marked Krea 3 out of beta on October 20, 2026, the same day Ultra mode reached the API and Krea published its benchmark results.

How does Krea 3 rank on image editing benchmarks?

As reported by Krea, Krea 3 Ultra ranks fourth on Arena's multi-image edit board, fifth on Artificial Analysis Image Editing, and fourth of 51 on Design Arena's image editing board. The boards use different scales, so compare ranks, not raw scores.

When should I use fast, high or ultra in Krea 3?

Fast for most edits and quick exploration, high for complex edits with many changes, and Ultra for final composites where several references and a product that can't change all have to come together. Krea's own data shows fast beating Ultra on broad style edits, so don't default to the slowest mode.