Introduction

Krea 3 just graduated. Krea's changelog for October 19 opened the Krea 3 API to every workspace, and the October 20 entry took the model out of beta and added Ultra thinking mode to the API. Krea 3 is Krea's in-house, closed image editing model that also does text-to-image. You can run it today in the Krea app (it's the default image model), through the Krea API, in the Krea iOS app, in Krea for ChatGPT, and in ComfyUI through Partner Nodes.

The same day, Ultra showed up on three independent image-editing leaderboards and landed in the top five on all of them. That's a big moment for a model that was behind a beta code last week ✨

What changed since launch

When we covered the Krea 3 API early access on October 15, you needed a beta invite or a sales email to touch the API. Here's what moved in five days, per Krea's own changelog:

  • October 16: Krea 3 became the default image model in the app for everyone, with Ultra on Max plans and above. API output got 1K and 2K resolution options and a priority setting.
  • October 19: the Krea 3 API went public. The fast and high thinking modes no longer need beta access.
  • October 20: Krea 3 left beta, thinking_mode: "ultra" arrived in the API, and Krea 3 showed up in the iOS model picker and the ChatGPT app.

If you already live in Comfy, our Krea 3 Partner Nodes piece covers the hosted Comfy route.

How Ultra works

Krea 3 has three thinking modes, and they change how much the model plans before it draws:

  • fast (default): quickly expands your prompt, then generates. Krea says it makes 2K edits in under 8 seconds.
  • high: a stronger reasoning engine plans multi-step or multi-reference edits. Slower, and it adds $0.02.
  • ultra: generates several candidates, reviews them against your prompt and references, and returns the best one. It always outputs at 2K, and Krea says a job can take up to about 10 minutes (about 16 minutes on low priority).

Ultra comes with some rules, so read these before you build it into a pipeline:

  • It needs a Max plan or above. Other workspaces get a 402 error with K3_ULTRA_ACCESS_REQUIRED.
  • It needs at least one reference image. A text-only Ultra request quietly runs as fast and is billed as fast.
  • It does not support regional_edit, negative_prompt, width/height, LoRAs, style references, or Zero Data Retention.
  • Each request returns one image, so use a webhook or poll with a long timeout.

Where it ranks

Three leaderboards added Krea 3 Ultra on October 20:

  • Artificial Analysis (AA-Image-Editing v2.1): Ultra is #5, with quality AA calls similar to Nano Banana 2.1 and MAI-Image-2.6. AA says it's the highest-ranked editing model from outside OpenAI, Google and Microsoft, and it's top 7 in 15 of AA's 17 editing categories, including #2 in Animation & Gaming and #2 in Frontier. The default Fast mode places #11.
  • Design Arena (Image Editing): Ultra is #4 of 51, ahead of Reve 2.0, with #2 in Architecture and #3 in Landscape.
  • LMArena: Ultra is #4 in Multi-Image Edit with 1435 points (#2 in Commercial Design and in 3D) and #8 in single Image Edit with 1411 points.

Krea's own summary is that Ultra is beaten only by OpenAI models in these arenas. The arena numbers are a snapshot from launch day and will move as votes come in.

Pricing

Krea 3 is billed per image, and the price stacks by resolution, references, and thinking mode:

  • Base price: $0.03 at 1K output, $0.06 at 2K output.
  • Each 1K reference image: +$0.02.
  • Each 2K reference image: +$0.07.
  • high mode: +$0.02 per generation.
  • ultra mode: +$0.20 per generation, at 2K only.

So a 2K Ultra edit with one 1K reference adds up to $0.28 by the docs table. Artificial Analysis quoted $0.27 for the same setup, so check your own invoice if pennies matter at volume.

Try an Ultra edit

Here's the shape of an Ultra call, based on the example in Krea's API overview. Upload your references with POST /assets first, or pass public image URLs. Then submit and poll GET /jobs/{job_id} until the status reads completed.

curl -s https://api.krea.ai/generate/image/krea/krea-3 \
  -H "Authorization: Bearer $KREA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Put the sneaker from the second image on the person in the first image",
    "image_urls": ["PERSON_IMAGE_URL", "SNEAKER_IMAGE_URL"],
    "thinking_mode": "ultra",
    "aspect_ratio": "4:5"
  }'

Krea's prompting advice is short and good: lead with the change, add only the details that help, and close with one preservation sentence. When you use more than one image, refer to each by what it shows. A paste-ready edit prompt in that style:

Replace the café table with a weathered oak workbench covered in brass tools and loose sketches, and warm the window light to late golden hour, keeping the person, their pose and the framing exactly the same.

For quick everyday edits, start on fast with "regional_edit": true so everything outside the change stays pixel-exact. Save Ultra for the final pass.

Weights and what's next

Krea 3 is still closed, but Krea now says it is the only frontier model that supports licensing of its weights, for custom fine-tunes or cheaper self-hosted inference. That's a sales conversation through Krea's site, not a download. Krea also says LoRA training support and Style References are coming. Right now the API answers both loras and style_urls with a "coming soon" error, so don't plan on them yet.

Original Source

Krea's launch-day post:

Krea's changelog lists the October 19 and 20 updates, and the Krea 3 API overview has the full parameter list, Ultra rules and pricing table.

Conclusion

Five days ago Krea 3 was a gated beta. Now any workspace can call it, the app uses it by default, and the Ultra mode is trading places with the biggest labs on editing leaderboards. If you edit images for a living, run your hardest edit through Ultra once and see if the extra twenty cents earns its keep 💛

—Aurelia ♡