# Krea 3 Beta: Image Editing, API Pricing and What to Check

Krea's new image model puts fast, controlled edits first, but its beta access and pricing need to be read separately for each provider.

By Paras Madan | 2026-10-16


Krea opened the Krea 3 beta on October 15, 2026, introducing an image-editing model with bounding-box control, reference images and fast repeat edits. The useful question for creative teams is whether changing one detail can leave the approved parts of an image alone.

The [official announcement thread](https://x.com/krea_ai/status/2110760037128012146) claims end-to-end 2K edits in under eight seconds. Krea says it is [available to everyone](https://x.com/krea_ai/status/2110760046934302958), while the [final release and more benchmarks](https://x.com/krea_ai/status/2110760045214625863) are planned for the following week. Those are different stages of availability, so this should still be evaluated as a beta.

## What Krea 3 changes in an editing workflow

Krea's [launch page](https://www.krea.ai/krea-3) presents regional editing: draw a box, describe a change and preserve the surrounding image. It also advertises up to eight references, native 2K output and control over product colours using hex codes.

These controls address a familiar problem with generative editing. A request to change a product's colour can also change its shape, label or lighting. The result may look good while becoming less useful for the original brief.

Consider a product image whose composition has already been approved. A designer might want a second colour variant while retaining the packaging text, camera angle and background. A bounded edit gives that request a more specific target than asking the model to reinterpret the entire scene.

**The practical test is how much approved work survives the next edit.** Compare the label, edges and shadows with the original at full resolution. Then make another change and compare again. A convincing first result does not establish that a sequence of edits will remain consistent.

## Krea 3 pricing: the website and API use different descriptions

Krea's launch page advertises roughly $0.06 per image. Developers should check the provider's actual billing meter before turning that headline into a budget.

The current [fal image-editing endpoint](https://fal.ai/models/krea/v3/edit) lists $0.025 per megapixel. It counts both the output and each reference image as processed by the model, with references larger than the output counted at the output's size. fal defines one megapixel as 1024 × 1024 pixels.

Under that published rule, a 1024 × 1024 output plus one reference processed at the same size would cost $0.05. Three references of that size plus the output would cost $0.10. These are calculated examples, not measured invoices, and changing resolution changes the calculation.

That makes reference selection an economic decision as well as a creative one. Adding every available angle may cost more without helping the particular edit. Use the references needed to specify the object and the requested change, then judge whether the additional context improves the result.

## Krea 3 API access still needs a provider check

fal's current [API schema for `krea/v3/edit`](https://fal.ai/models/krea/v3/edit/api#schema-input) accepts a prompt and one to eight reference images through `image_urls`. It also exposes output-size and seed controls, plus two model variants: `berry` favours quality and `jam` favours speed.

The default `image_size` is `auto`, which lets the prompt expander choose an aspect ratio at roughly 1024 × 1024 pixels. Native 2K capability does not mean every request defaults to 2K. Set the output dimensions deliberately when comparing quality and cost.

The published input schema has no dedicated bounding-box or mask field. Do not assume that the regional controls shown in Krea's app are exposed through this fal endpoint in the same form.

fal's [broader Krea 3 introduction](https://fal.ai/krea-3), marked as updated on October 13, still describes early access, a four-reference limit and an earlier endpoint name. Use the current endpoint documentation for integration details, and verify the access and quota available in your account. No inference request was run for this article.

## How to read the speed and quality claims

Krea compares its quality with Nano Banana Pro and GPT Image 2.5. Those are vendor claims at this stage, rather than a settled ranking from independent testing.

There is also a useful qualification on the launch page's comparison animation. It uses pre-generated outputs, with an illustrative eight-second playback time for Krea 3. It is not a live benchmark. The page describes competitor timings from a twelve-image batch and labels one competitor price as a draft estimate.

A team evaluating the model should use its own repeated edits, including failures, and measure the time until an acceptable result is ready. Network delays, retries and manual corrections all belong in that measurement.

For product photography and design review, faster generation becomes valuable when it lets people test a specific change while the discussion is still happening. The next thing to watch is whether Krea 3's final release makes that control dependable across ordinary, imperfect source images.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn model launches and product updates into clear stories that developers and founders can understand and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).