Krea 3 Lands at $0.06 an Edit, and Puts a Stopwatch on Image Editing

The interesting number in Krea 3 is not the price. It is the stopwatch. Krea released Krea 3 on 15 October 2026 as an image editing model, and the headline it chose for itself is a latency claim: most edits come back in under 8 seconds, end to end, at native 2K, for $0.06 per image. Against that, the company's own comparison chart puts Google's Nano Banana Pro at 33 seconds and $0.135 per image, and OpenAI's GPT Image 2 at 90 seconds and $0.22. Those two figures are Krea's, measured on Krea's own traffic, and they are the reason a model launch that would normally be a footnote about a new checkpoint is worth reading closely — because for the first time in this generation of editors, the price of a catalogue job is set by seconds rather than by cents.

Everything in this piece that comes from Krea is labelled as Krea's. There is no independent benchmark of Krea 3 yet, and no third party has published a measured edit time for it. That gap is the whole story underneath the story.

What actually shipped on 15 October

Krea 3 is an editing model, not a generator with an edit mode bolted on. Krea's own FAQ is blunt about it: the model is "trained for editing," it is not a distilled copy of something larger, and it spends its compute on the region you change rather than redrawing the whole frame. The vendor says that is where the speed comes from — find the region first, render only what changes — and it is the same mechanism behind the accuracy claim.

The published capability list, as the vendor states it:

• Native 2K output, with Krea's own examples quoted at 2736 × 1536 pixels. 1K is also offered at half the price.

• Up to 6 reference images per edit, per the launch post. The product page separately advertises "up to eight references" with a numbered-image prompt style for pointing at people, products, places and styles. The two numbers sit on two pages of the same site and Krea has not reconciled them.

• Region-scoped editing. Draw a box, describe what changes, and the rest of the frame is supposed to stay exactly as it was — including, Krea stresses, product labels and logos.

• Character consistency across a sequence of edits, and style references on top.

• Voice-driven editing in the consumer app: send a photo, then a voice note, and the edit comes back before you finish the sentence.

• 251 launch presets, billed as day-to-night, blockout, sketch, re-skin, new sky and green screen.

The consumer pricing page is where the naming gets slightly loose: the same model appears in Krea's plan tables as Krea 3 Beta, while the launch page and the blog call it plain Krea 3. Treat the identifier you write into a config as something to verify rather than assume.

There is one more figure that the launch page does not lead with and the developer documentation does. The $0.06 is a base price for a 2K output with no references attached. Krea's API reference prices references separately: each 1K reference image adds $0.02 and each 2K reference image adds $0.07, on top of the base, at either output size. Running the edit in the slower "high" thinking mode adds another $0.02 per generation. A 2K edit carrying four 2K references therefore costs $0.06 plus $0.28, or $0.34 — a number worth having before you size a catalogue job on the headline rate. Krea's own API documentation also puts access in early access, callable only from workspaces inside the Krea 3 beta, with access requested rather than self-served.

The speed chart, read carefully

Krea publishes its editor comparison twice, and the two publishings do not agree.

The launch blog's comparison table gives Nano Banana Pro 33 seconds and GPT Image 2 90 seconds, with a footnote that these are median times from request to finished image for edit jobs run on Krea between 24 September and 1 October 2026, queue time included. The product page's side-by-side animation uses a different pair of numbers again — 26.8 seconds for Nano Banana Pro and 39.9 seconds for GPT Image 2.5 Sunburst — and describes them as the median request-to-result time of each model's 12-image batch. A separate line on the same page says GPT Image 2.5 pricing is "a draft estimate pending validation."

That is not a scandal, but it is a disclosure you should hold on to. Measuring a competitor's latency by routing your own traffic to it tells you what that model does from your infrastructure, on your prompts, at that hour — a real and useful number that is not the same thing as an independent latency benchmark. The spread between 39.9 and 90 seconds for OpenAI's editors is the size of the uncertainty in the comparison, and it is wider than the gap Krea is advertising.

What is stable across both pages is the claim about Krea 3 itself: under 8 seconds, end to end, at 2K, $0.06 per image. That is the number to test first, because it is the one the whole pitch rests on.

The precision claim, and how it is measured

The second half of the pitch is what the model does not touch. Krea defines precision as the share of pixels outside the edit area left unchanged, compared at quarter resolution after best alignment so that a difference in output size does not count as a change, with one attempt per model at 2K on the same image and the same prompt, and Krea 3 running in a "Precise" mode for the test.

Two things are worth saying about that. First, it is a self-run comparison — Krea chose the prompt, the image and the mode. Second, the metric itself is the right one for the problem. The failure that ruins an edit in production is not a bad sofa; it is a good sofa that came back with a subtly restyled rug, a different window view and a logo that now reads slightly wrong. Any model that scores well on unchanged-pixels-outside-the-box is solving the thing practitioners actually complain about, regardless of who ran the test.

Krea's own worked examples lean on exactly that: swapping the seating in a room while the rug, the view and the reader in the corner stay put; turning a technical line drawing plus a six-colour palette into an architectural photograph with each colour where the brief put it; putting the same cast on a new street under new light.

What $0.06 does to a catalogue job

Image editing pricing has been quoted per image for a while, so the interesting move here is not the unit — it is the ratio. Krea's own volume arithmetic, presented on its launch page, compares a 100,000-image month at $0.06 against the same month at $0.135 and arrives at $6,000 versus $13,500. Treat both figures as vendor working, because they are. The shape of the argument survives the caveat anyway: at $0.135 an edit is a considered action you schedule, and at $0.06 an edit is something you can run five times and pick the best one, which changes the workflow rather than the invoice.

It also changes what a retry costs. If a first pass misses on the label, spending another six cents is not a decision anyone has to make. If it costs twenty-two, somebody has to.

What is not in the release

Three absences define the limits of what can be claimed today.

There are no open weights. Krea's own FAQ says self-hosting is available on enterprise plans, running on your own GPUs in your own cloud, which is a licensing conversation rather than a download. Nothing about Krea 3 is in a public repository.

There is no independent benchmark. No third party has published an Elo, a head-to-head or a measured latency for Krea 3. Every performance figure in this article traces back to Krea, and the comparison figures for the competing models were gathered by Krea as well.

There is no published architecture. Krea says the model is new and trained for editing rather than distilled, and that it renders only the changed region — but it has published no parameter count, no paper and no training detail, so the mechanism behind the speed claim is a description rather than a specification.

Against that, there is one piece of real third-party signal in the release, and it is a customer rather than a benchmark: Pixelcut's co-founder and CEO is quoted saying his team had early checkpoints in front of paying users, and that what stood out was lighting on product shots. That is a partner testimonial, not an evaluation, and it should be read as one — but it is a claim from someone who put the model in front of customers before launch, which is more than a leaderboard row usually tells you.

Where you can call it, and where this blog sits

Krea 3 is live in Krea's own app and API, and the vendor says partner platforms serve it as well. If you are choosing an access route rather than a model, the useful question is whether the editors you already run sit behind the same key as the editor you are about to try, because an editing pipeline that requires a separate contract per model is a pipeline nobody experiments in.

One honest note about us: OrcaRouter does not route Krea 3 today, so calling it means calling Krea. What we do provide is one OpenAI-compatible endpoint over a few hundred models, with provider list prices passed through at no markup and automatic failover when a provider errors or rate-limits. The reason that matters here is narrower than a sales pitch: the models Krea 3 is being measured against — GPT Image 2, GPT-Image-2.5 Flare and Sunburst, and Grok Imagine Image — are the kind of thing you want to A/B on your own images before you accept anyone's median, and running that test through one endpoint is a configuration change rather than four integrations.

What to watch next

The first thing to watch is an independent time-to-edit measurement, because the entire positioning is a latency claim and the only stopwatches on it so far belong to the vendor. The second is whether the 6-versus-8 reference discrepancy and the Krea-3-versus-Krea-3-Beta naming get reconciled, since both are the sort of small inconsistency that turns into a support ticket when you build against them. The third is whether a third party runs the precision test with its own prompt, because "only what you asked" is a measurable claim and the industry has not yet measured it for anyone.

Until then, the fair reading of Krea 3 is that it is a shipped, priced, immediately usable editing model with a credible mechanism behind its speed claim and no external verification of any of it. At $0.06 an edit, the cost of finding out for yourself is low enough that waiting for a leaderboard would be the more expensive choice.