Krea 3: Under 8 Seconds an Edit, $0.06 an Image, and a Reference Count Krea Can't Agree On

Krea released Krea 3 on 15 October 2026, and the pitch is a trade rather than a capability. Its launch post opens by conceding that editing models got good and then got slow and expensive, and offers Krea 3 as the other side of that trade: most edits finish in under 8 seconds end to end at native 2K, for $0.06 an image. The company's own comparison chart puts Nano Banana Pro at 33 seconds and $0.135 per image and GPT Image 2 at 90 seconds and $0.22, measured on Krea's own edit traffic between 24 September and 1 October 2026. That is a vendor talking about competitors using vendor-collected medians, and it is worth reading as such — but the underlying claim, that an edit now costs less than a second of a designer's patience per cent spent, is the reason a small model launch is worth an hour of your time.

The two things to settle before you build anything on Krea 3 are how many reference images you are actually allowed and what each one costs. Krea's own pages answer the first question three different ways, and the second answer is a meter.

What Krea 3 actually is, and what it is not

Krea 3 is an editing model, not a generator with an edit mode bolted on. The API's own description says it "follows plain-language instructions: you say what should change and what should stay the same" — and it will also generate from text alone if you send no reference image, but generation is the side job. The product page's framing is narrower still: it edits, it changes only what you ask, it renders natively at 2K, which Krea illustrates with a 2736 × 1536 pixel example.

Three things in the API are the real product, and none of them is the price:

• Regional editing. A regional_edit/ flag regenerates only the parts of the frame that need to change and leaves the rest pixel-exact. Krea's documentation recommends it for chained edits — recolouring or swapping an object, fixing a detail, changing text on a sign — and warns that ordinary (whole-image) edits accumulate artifacts when you keep editing the model's own output.

• Two thinking modes. thinking_mode/ is either fast/, the default, which expands the prompt and generates, or high/, which spends a stronger reasoning pass planning multi-reference and multi-step edits before generating. Krea's guidance is to use high/ when a fast result missed the request.

• Reference composition. The capability Krea leads with socially: give it a face, an outfit and a room as separate inputs and it assembles one scene, keeping the person's identity and the grading steady. Krea's blog demonstrates it with a prompt that reads like a casting sheet — reproduce the four characters from the references exactly, change only their poses.

Not in the box: open weights. Krea 2, published on 12 May 2026, shipped open checkpoints and a community licence. Krea 3 does not, and self-hosting is an enterprise conversation. Also not in the box, as of 16 October 2026: a single independent score. Krea 3 appears on no public image-editing leaderboard, which is unsurprising for a model that is one day old and gated behind a beta workspace.

The reference limit is three different numbers on three Krea pages

This is the finding that changes how you'd budget the model, and it does not require a benchmark to establish — only reading the pages in one sitting.

• The launch post and its FAQ: "up to 6 reference images per edit."

• The product page: "Up to eight references. One image. Point at people, products, places and styles by number."

• The API reference: the image_urls/ array takes "up to 8" — but with five or more references in one request, each reference is resized to at most 1536 px on its longest side, and the docs tell you plainly that "for full-detail references, send four or fewer."

Those are three different products. A designer reading the product page plans a shoot around eight references and their fine detail. An engineer reading the API reference learns that reference five onward is a downscaled reference, and that the number to design around is actually four. And a team that needs Zero Data Retention — Krea 3 supports it for enterprise workspaces via an X-Api-Zero-Data-Retention: 1/ header — gets a hard cap of 4 references per request, because in that mode every reference must be sent as a base64 data URI.

The interaction with identity is the part that should slow you down. The whole point of the character/outfit/location split is that the face survives the composition. The five-plus reference path is precisely the path where Krea's own docs resize the face you were trying to preserve.

What a Krea 3 edit costs, worked through

The API reference publishes a rate card that is per component rather than per image, and the components stack:

• Base price — $0.03 for a 1K output, $0.06 for a 2K output.

• Each reference image — +$0.02 if the reference is 1K, +$0.07 if it is 2K, charged at either output size.

• high/ thinking mode — +$0.02 per generation.

Work that forward and the headline inverts fast. One reference at 2K is $0.13. Four references at 2K — the count Krea's own docs call the full-detail ceiling — is $0.34. Six references at 2K, the number on the launch post, is $0.48, and $0.50 in high thinking mode. The "$0.06 per image" figure everybody will quote is the price of the simplest possible edit: one 2K output with nothing attached to it.

At catalogue scale it is still cheap against Krea's own comparison — Krea's chart puts 100,000 Krea 3 edits at 2K at $6,000 per month against $13,500 for Nano Banana Pro and $22,000 for GPT Image 2 — but that chart is explicitly "list price per image, without image references." At six 2K references per edit, the same 100,000 edits land near $48,000. Whether the reference meter is 40% or 800% of your bill depends entirely on how many images your pipeline attaches, and that is a question about your workflow, not about the model.

The billing is also silent on one point that the docs raise. A reference sent as the fifth or later is resized to at most 1536 px. The rate card charges $0.07 for a 2K reference. Krea does not state whether a reference the model itself downscaled still bills at the 2K rate. On a six-reference edit that ambiguity is $0.14 per call — a quarter of the bill — and it is worth an email before you commit a budget.

The speed claim, and the footnote under it

Under 8 seconds end to end, queue included, is the claim Krea leads with, and the footnote attached to it is unusually specific: the competitor timings come from median request-to-finished-image times for edit jobs run on Krea's own platform between 24 September and 1 October 2026. That is not a neutral benchmark and Krea does not pretend it is — the same chart says the prices are "list prices per image, without image references," which is exactly the qualification that matters for Krea 3 itself.

The honest state of the number: no third party has reproduced it, and Krea's own materials contain at least one other latency figure for the same model on a different page, gathered against a different set of competitors. Measure it on your prompts at your reference count, because the reference count is the variable that changes both the seconds and the cents.

CPI-Bench: the launch number we could not find

Coverage of the Krea 3 launch has circulated a CPI-Bench score for the model — a figure in the 4.8-out-of-5 range, presented as a real-world image-editing benchmark. We could not find it. It is not on the launch post, not on the product page, not in the API reference, and a defunct CPI-Bench leaderboard is not returned by any search engine we can reach from here. That is not proof it does not exist; Krea's product page has an FAQ section that renders client-side and the model was previewed publicly in late August 2026, so the number may well live somewhere we cannot load.

What it means for you is simpler than the mystery: do not put a CPI-Bench number in a procurement document. The only figures about Krea 3 that anyone can check today are Krea's own published price schedule, Krea's own reference limits and Krea's own measured timings. Everything else about this model is currently unaudited by anyone but its maker — and that includes the composition claim that is the actual reason to use it.

Where a router fits, and where it does not

OrcaRouter does not route Krea 3. It is not in our catalogue. Krea 3 is called at Krea's own API, which is in early access and requires a workspace in the Krea 3 beta; the vendor's app and API serve it, along with several third-party platforms Krea names on its launch page. Nothing in this piece asks you to believe otherwise.

What that leaves you is an integration question with a familiar shape. A beta-gated endpoint with three published reference limits and a per-reference meter is the right place to put work you can afford to re-run, not work you have promised to a client on Friday. The models you have already measured belong on a single endpoint covering 200-plus models with no markup on provider list prices, so that the day Krea 3 leaves beta is a routing change rather than a new contract, a new SDK and a new billing relationship. Until then, automatic failover that moves a failed call to a backup model instead of into your job queue is what keeps a beta's bad afternoons from becoming your incident report.

Krea 3 is the most interesting editing model of the week and possibly the most awkward one to plan around. The capability — separate references for a person, a garment and a place, composed into one frame with the identity held — is the part that is genuinely new, and it is exactly the part that Krea's five-reference resize rule and four-reference Zero Data Retention cap cut into. Read the reference limit before the price, and read your own workflow before both. Then run one real catalogue job through it, count the references, and see whether the bill looks like $0.06 or like $0.48.

That gap between what a scoreboard measures and what a deployment meets is the reason the record matters: the documented AI failures and agent incidents around models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.