Introduction

The Krea 3 wait just got a lot more concrete. Krea's official changelog, dated October 15, 2026, now lists Krea 3 API in early access, along with a full developer overview: endpoint, parameters, thinking modes, regional edits, and a price sheet. Krea 3 is a closed, API-and-app image editing model that also does text-to-image. Krea's own phrasing: say what should change and what should stay the same, and it returns the edit in seconds.

One honest caveat up front: this is not general availability yet. The API is open only to workspaces in the Krea 3 beta (access requests go through sales@krea.ai), the in-app model is labelled Krea 3 Beta behind an access-code page, and fal still lists it as request-access early access. But the spec is public, the pricing is public, and that's what makers need to plan around.

What shipped today

From Krea's changelog and the new Krea 3 API overview:

  • Endpoint: POST /generate/image/krea/krea-3 on api.krea.ai, using Krea's standard job lifecycle (you get a job_id, then poll or use a webhook)
  • Up to 8 reference images per request via image_urls (asset uploads, external URLs, or base64 data URIs). Leave them out and it does text-to-image
  • Two thinking modes: fast (the default, which quickly expands your prompt then generates) and high (a stronger reasoning engine that plans complex, multi-step, or multi-reference edits; slower)
  • Regional edits: set regional_edit: true and Krea 3 regenerates only what needs to change. Everything else stays pixel-exact, so you can chain edits without quality loss
  • Output: 1k or 2k, with aspect ratios from 1:1 through 21:9 (including 3:2, 9:16, and 16:9), or custom width and height up to 4096 that get rounded to the nearest supported canvas
  • Extras: seed, plus a negative_prompt up to 1,024 characters
  • Zero Data Retention for enterprise workspaces, which caps you at 4 references and requires base64 inputs

There's also a housekeeping note for anyone already on the beta: the old thinking_mode values instant and medium become fast and high, and the legacy names stop working on October 23, 2026.

Pricing

Krea prices each image by output size, thinking mode, and references:

  • Base: $0.03 per 1K output, $0.06 per 2K output
  • Each 1K reference image: +$0.02
  • Each 2K reference image: +$0.07
  • High thinking mode: +$0.02 per generation

So a plain 1K text-to-image call costs $0.03, and a 2K edit with one 1K reference in high mode works out to $0.10. Edit-heavy pipelines will feel the reference fees most, so it's worth taking Krea's own advice below about reference sizes.

How Krea says to prompt it

Krea 3 is pitched as an instruction-following editor, and the docs are refreshingly specific:

  • Lead with the change, and add only the details that help execute it
  • Close with one preservation sentence instead of listing everything that should stay the same
  • With several images, refer to each by what it shows ("the sneaker from the second image")
  • Around 1 MP references are enough for most edits; references over 4 MP get resized to 4 MP
  • With five or more references, each is capped at 1536 px on the long side, so send four or fewer when fine detail (small text, logos, fabric) matters
  • Use regional edits for local fixes like recolors, swaps, or sign text. Skip them for global restyles, relighting, or season changes, where a full re-render is the point

A paste-ready edit prompt in that shape:

Swap the paper coffee cup in the first image for the matte black ceramic mug from the second image, matching the warm window light and soft shadow on the table. Keep everything else in the scene exactly the same.

Where to run it today

  • Krea app: listed as Krea 3 Beta on Krea's models page. Access is code-gated through Krea's beta access page, and Krea handed out early codes on X on October 14 (US time)
  • Krea API: beta workspaces only for now; request access via sales@krea.ai
  • fal: early access through fal-ai/krea-3 (edit with 1 to 4 references, plus text-to-image), request access on fal's Krea 3 page

If you're in the beta, a minimal request body looks like this:

curl https://api.krea.ai/generate/image/krea/krea-3 \
  -H "Authorization: Bearer $KREA_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Place the product on a clean studio background. Keep the product exactly the same.",
    "thinking_mode": "fast",
    "image_urls": ["https://example.com/your-product.png"],
    "resolution": "2k",
    "regional_edit": false
  }'

Then poll GET /jobs/{job_id} until status is completed and grab result.urls[0].

Original Source

Krea's changelog entry and API overview are the primary sources: Krea changelog and Krea 3 API overview. Here is Krea's pre-launch post with the early access codes:

Conclusion

Krea 3 is shaping up as a direct shot at the instruction-editing crowd: plain-language edits, up to eight references, and a regional mode that promises pixel-exact chaining, all starting at three cents an image. It's still behind a beta gate today, but the docs are live, so you can design your pipeline now and be ready the moment the doors open wider. We'll keep watching for the full public launch.

—Aurelia ♡