Ideogram Vector 1.0 vs Krea 3: Price per Attempt, Not per Picture

The per-thousand-outputs arithmetic produces a result almost nobody expects. Krea 3 at its Ultra setting is listed at $270 per thousand images on Artificial Analysis's price column. Ideogram Vector 1.0, the vector model Ideogram released on October 22, 2026, is $130 per thousand at its low quality tier, $160 at medium and $200 at high. The "cheap raster model versus expensive vector model" story that most 2026 design-tool roundups still tell is, at the tier people actually use, backwards. The expensive thing in this comparison is not the SVG. It is the model that generates and re-scores multiple candidates per request.

So price is a dead end, and so is quality, because no independent score exists for either side's output in a shared format. What is left is the question a studio actually asks when it budgets a job: how many attempts does it take to reach an asset I can hand over, and what does each attempt cost me? That is the axis this comparison uses, because it is the only one on which both models can be measured with numbers that are publicly checkable.

The two models, in one sentence each

Krea 3 is an instruction-based image editor that launched on October 15, 2026 and became Krea's default image model the next day; its API opened to all workspaces on October 19. It is built to take an existing image and change part of it — a product shot, a character, a region — in under eight seconds at native 2K, according to Krea's own measurement. Text-to-image is supported but secondary.

Ideogram Vector 1.0 is not an editor. It takes a prompt and emits an SVG, or takes a raster image and converts it into an SVG. There is no edit endpoint, no inpainting, no region selection. You get a file, and if the file is wrong you generate another one.

That asymmetry is the whole article. One model is optimised for the loop before the asset exists. The other is optimised for the artifact itself.

What a Krea 3 render actually costs

The $0.06 headline that Krea puts on a 2K image is real, but it is the price of a request with no reference images and no thinking mode. A production edit usually has both. The published components:

• Output price — Krea 3 is $0.03 for a 1K output and $0.06 for a 2K output, billed per completed job; failed or cancelled jobs are free.

• Reference price — each 1K reference image adds $0.02, and each 2K reference adds $0.07, charged at the output size. A 2K edit with one 2K reference is therefore $0.13, not $0.06. With four references it is $0.34.

• Thinking mode — high mode adds $0.02 per generation; the default fast mode does not.

• Reference limits — Krea's launch material says six references, its API reference says eight, and images five and above are downscaled to a 1536-pixel long side. Four or fewer is what the documentation recommends for full detail.

• Ultra — the setting that scores best is not a separate model but a mode that generates several candidates, scores them and returns the winner. It requires at least one reference image, outputs 2K only, drops regional editing, needs a Max plan or above, and can take up to about ten minutes.

• Measured quality — on Artificial Analysis's image-editing board, Krea 3 Ultra sits fifth at 1,152 Elo from 5,440 blind votes, inside a statistical tie with MAI Image 2.6 Preview and Nano Banana 2.1. Krea 3 Fast, which is the default API mode, sits eleventh at 1,126 Elo from 5,144 votes.

Those last two lines matter for budgeting. The mode that wins the benchmark is the mode that costs $270 per thousand and takes minutes. The mode that costs $60 per thousand is six ranks and 26 Elo lower. You are not choosing between Krea and Ideogram at a single price point; you are choosing between two Krea price points before you even get to the vector model.

The attempt budget

Here is the calculation that should replace the price-per-image comparison. Call it expected cost to a usable asset: the number of attempts multiplied by the price of an attempt, plus the human minutes spent reviewing each one.

A mark for a client might take five attempts before someone signs off. On Krea 3 at 2K with one 2K reference, five attempts cost $0.65, and the fifth attempt can be a surgical correction: regional editing regenerates only the flagged area and leaves the rest of the frame pixel-exact, which is exactly the mechanism that makes attempt six cheaper and more likely to succeed than attempt five. The loop converges.

On Ideogram Vector 1.0 at high quality, five attempts cost $1.00, and every attempt is a fresh generation from a fresh seed. There is no way to say "keep everything, move the counter left by four pixels." A rejected SVG teaches you something about the prompt and nothing about the file. The loop restarts.

That difference is worth more than the 35-cent gap suggests, but it cuts both ways. Krea's five attempts buy five raster drafts and no vector file. Vector 1.0's five attempts buy five SVGs, of which perhaps one is usable — but that one is already the deliverable. If the job ends in a raster asset, Krea wins on both cost and convergence. If the job ends in paths, Krea has not finished; it has produced an input to a tracing step.

Where the vector file ends the loop

Print and manufacturing workflows consume geometry. Embroidery machines, vinyl cutters, screen-print separations and laser etchers all want paths, and a raster file is an input to a trace in every one of them. The trace is where a clean mark becomes four hundred anchors along what should be a straight edge.

Ideogram Vector 1.0's entire proposition is that the trace does not have to happen. At $0.13 to $0.20 per SVG, or $0.32 at high quality with reference images, you get shapes, gradients and typography as editable objects. The company's claim is that the files are clean and layered. That claim is the one thing in this comparison that nobody outside Ideogram has checked: there is no published anchor count, no independent inspection of a real output file, and no benchmark for vector quality anywhere in the industry. The launch coverage repeats the marketing language and stops there.

Krea has no answer to this and does not pretend to. It is a raster editor, and a very good one at the thing it does.

The workflow that uses both

The rational production pattern is not a choice between these models. It is a handoff, and the handoff has a price worth calculating.

Explore in raster, because exploration is where attempt count explodes. Krea 3's text-to-image path at $0.03 for 1K and $0.06 for 2K lets you run twenty directions for under a dollar. Pick the one that works. Then send the chosen render through Ideogram Vector 1.0's vectorizer at $0.10 per image.

Total for a traced asset: about $0.25 if you explored at 1K, or $0.28 at 2K. Compare that with $0.13 for a direct low-quality SVG. Direct generation is cheaper — but only if the first or second SVG is usable, and there is no way to know that in advance. The hybrid is more predictable because the raster stage is where you spend your failures.

The hybrid carries one documented caveat. The vectorizer's layer-grouping option is described in Ideogram's API reference as requesting grouping of shapes into layers, and the reference states that grouping is not yet available. What comes back is geometry, but the layer organisation the marketing describes may not be there. Test on one file before you build a pipeline around it.

Running the pipeline rather than the model

Neither Krea 3 nor Ideogram Vector 1.0 is on OrcaRouter's catalogue today, and we will not pretend otherwise. The image call is also not where most of a design pipeline's complexity lives. The calls that multiply are the ones around it: a model that writes the brief, a model that turns the brief into forty prompt variants, a model that reviews the forty outputs against the brief and flags the six that failed.

That layer is what OrcaRouter covers: one API for more than 200 models with provider list prices passed through at 0% markup, automatic failover when a provider stalls in the middle of a batch, and a routing DSL that can put a prompt-writing model and a review model behind a single endpoint. For a studio running this loop weekly, the practical benefit is not the image price. It is that a stalled provider does not take the batch down with it, and that a second model is a config change rather than a second contract.

How to choose in one question

Does the deliverable have to be editable geometry? If yes, Krea 3 cannot finish the job and Vector 1.0 can — subject to the one unverified claim about file cleanliness. Start with the vectorizer on a render you already like, inspect the result, and only then decide whether direct SVG generation is worth the premium.

If the deliverable is a raster image that will live on a screen, in an app, or in a campaign, Krea 3 is the stronger and, at the fast tier, cheaper tool by a wide margin. Its reference composition, regional editing and 2K native output are built for exactly that work, and the Ultra tier buys measurable quality when the edit is hard enough to justify $0.27 an image.

If you need both, budget for both. The uncomfortable truth is that the only number that would settle Vector 1.0's side of this comparison — how many anchors a typical generated SVG actually contains — has not been published by Ideogram or measured by anyone else. Until it is, the vector model is a bet on a vendor sentence, and the responsible move is to ask for one sample file and open it in Illustrator before you commit a print run to it.