Krea 3 vs Ideogram 4.5: Two Ways to Sell an Edit

Krea 3 and Ideogram 4.5 were both pitched as editing models in the last two weeks, and they answer the same question in opposite ways. Ideogram 4.5, live since 30 September 2026, returns your image at its own size with the untouched pixels copied across, and charges by how many pixels it had to resolve — a ladder that runs from eight tenths of a cent to twenty-two cents per image. Krea 3, released on 15 October 2026, finds the region it needs to change and renders only that, which is why Krea quotes a flat $0.06 per image at 2K and under eight seconds end to end. One sells resolution fidelity; the other sells throughput. That difference decides which of them you can afford to run over a catalogue, and it is not the difference either company leads with. automatic failover provider list price passed through at 0% markup

The bill arrives before the benchmark does

Take a hundred thousand edits, which is a modest catalogue refresh, and put list prices next to each other.

• Krea 3 — $0.03 per image at 1K, $0.06 at 2K. Those are Krea's own published list prices on the launch page and in its FAQ, and the FAQ answers the obvious follow-up directly: "The launch price is the list price. We expect it to go down, not up."

• Ideogram 4.5 — a five-rung ladder, $0.008 at Very Low through $0.03 at Low, $0.06 at Medium, $0.10 at High, and $0.22 at High with a source image. Ideogram's own API pricing page is client-rendered and returned no figures to a text fetch, so the ladder comes from launch coverage rather than from the vendor's page. The top rung is corroborated independently: Artificial Analysis lists Ideogram 4.5 editing at $220.00 per 1,000 images, which is $0.22 an image to the cent.

• At volume — 100,000 edits at 2K, Krea 3 is $6,000 a month. The same 100,000 edits at Ideogram 4.5's High-with-source rung is $22,000. Krea's launch post prints that $6,000-versus-$22,000 comparison against GPT Image 2 rather than Ideogram; the arithmetic is identical because both top out at $0.22.

The honest caveat is that the two ladders are not measuring the same thing. Ideogram bills by resolved output pixels — up to 1024×1024 counts as 1K, anything above as 2K — so its cheap rungs are genuinely cheap for small outputs and its expensive rung is for a full-quality image-to-image pass. The Medium rung at $0.06 is a straight tie with Krea 3 at 2K, and comparing Krea 3 against Ideogram 4.5's High tier alone would flatter Krea. What survives the caveat is the shape: Krea 3 publishes one price and holds it across resolutions that matter, while Ideogram publishes a range and the useful quality settings sit at the top of it.

Where the pixels go, and what it costs to leave them alone

Krea's central claim is a negative one: it changes only what you asked. The launch page demonstrates it with a furniture swap — replace the sofa, the chairs and the table, and the rug, the view and the woman reading in the corner come back untouched — and with a precision figure it computes by counting pixels changed outside the edit area. Krea's own number for that is a share of pixels, not an Elo, and it is measured with Krea 3 in its Precise mode against Nano Banana Pro and GPT Image 2.5 Sunburst, one attempt per model, same image, same prompt. It is a vendor measurement of a vendor configuration, and it should be read that way.

Ideogram takes the opposite route to the same goal. Its Precise Edit endpoint returns the image at the source's exact width and height and copies untouched pixels rather than regenerating them, and it has been demonstrated on a 4,016 × 6,016 source — 24.2 megapixels — without downscaling. Krea 3 renders natively at 2K with 4K support for print; its precision story is about not needing to touch the rest of the frame, not about carrying a 24-megapixel plate through.

• Edit method — Krea 3 finds the region and renders only what changes vs Ideogram 4.5 copies untouched pixels and regenerates inside a mask or a natural-language instruction.

• Output sizing — Krea 3 at native 2K, 4K for print vs Ideogram 4.5 at the source image's exact dimensions, with generation sizes up to 2048 × 2048 unless a source is supplied.

• References — Krea 3 takes up to six per edit vs Ideogram 4.5 up to four on the edit endpoint, three of those if a mask is present.

• Instruction control — Krea 3 adds bounding boxes for placing new elements and multi-turn edits that hold across turns vs Ideogram 4.5's JSON prompts with per-element text, description, colour palette and bounding box, up to 16 hex colours per image and five per element.

What each one claims and what neither has proved

Krea 3's headline numbers are Krea's own. Under eight seconds end to end at 2K, $0.06 at 2K, six references, trainable LoRAs. The footer on its speed demo is unusually forthcoming about the method: the timings come from the median request-to-result time of each model's twelve-image batch including queue time, measured on Krea's own platform between 24 September and 1 October 2026, and one of the comparison prices is described in the page's own text as "a draft estimate pending validation." That is a vendor benchmark with its methodology printed on it. It is still a vendor benchmark.

Ideogram 4.5's quality numbers are not vendor numbers at all, which cuts the other way. Ideogram published no benchmark for 4.5; the model is demonstrated, not scored. The scored figures come from Artificial Analysis, read 16 October 2026: image editing 1,063 Elo from 2,728 votes, at $220 per 1,000 images, and text-to-image 1,019 Elo from 2,421 samples at $100 per 1,000. For context on that board, the leaders are OpenAI's GPT Image 2.5 pair at 1,198 and 1,177, and the editing board's own top rung is 135 Elo above Ideogram 4.5.

Krea 3 has no Artificial Analysis entry at all. It launched the day before this article's research was collected, which is the whole explanation, and no amount of adjacent evidence substitutes for it. So the two models cannot be scored against each other today. What can be compared is everything above the score line: the price per edit, the output resolution, the reference count, the instruction interface, and the licence.

Licence, training and the thing that is genuinely new

Both are closed. Ideogram 4.5 has no checkpoint, no repository and no licence for the 4.5 generation; the open-weights programme covers Ideogram 4.0 only, and 4.0's self-serve commercial tier is $300 a month for 10,000 images of self-hosted quantised weights, explicitly not full precision. Krea has released open weights before — Krea 2 RAW and Turbo are on the Hub — but Krea 3 is not among them. Krea's FAQ puts self-hosting under enterprise plans, on your own GPUs, with licence terms and hardware sizing to be discussed.

The capability that has no counterpart on Ideogram's side is LoRA training on the editing model itself. Krea's claim is that Krea 3 is the first editing model you can fine-tune on a specific product, person or style, so a brand's packshot or a recurring character stays the same across a campaign. Whether "first" holds up is a claim about the field rather than a measurement, and it is Krea's claim. What is checkable is that the mechanism exists on Krea's side — the API reference carries a "Train a custom style (LoRA)" endpoint — and that Ideogram 4.5's documented instruction surface is prompts, JSON and masks, with no fine-tuning path published for the 4.5 generation.

Neither one is on our own catalogue

Worth saying plainly, since this is our blog. OrcaRouter does not route Krea 3 and does not route Ideogram 4.5. Our image line today is built from OpenAI's GPT-Image models, xAI's Grok Imagine image model, and Google's Imagen and Gemini image families — that is the list, and neither of these two models is on it. The reason the comparison is still worth running on our pages is the reason it is worth running at all: if you are choosing an editor for a production pipeline, you are choosing it for the next two years, and Krea's own FAQ says the launch price is the list price and expects it to come down.

Where a router does change the arithmetic is at the layer above the model. A pipeline that has both a generation step and an editing step is two vendor integrations, two keys, two sets of rate limits and two failure modes, and that is the part a single endpoint for 200-plus models with automatic failover and provider list price passed through at 0% markup removes. If a vendor cuts a price, the cut is live here the same day rather than whenever the integration is refreshed. If a model is deprecated or a provider degrades, the failover path is a routing decision rather than an incident.

Which one to pick

The decision falls out of the volume, not the demo reel.

• High-volume catalogue work at 2K — Krea 3. A flat $0.06 and a sub-eight-second round trip is a different kind of product from a five-rung ladder whose useful settings start at $0.10. The launch price is stated as durable and the vendor says it expects it to fall.

• Large plates that must come back exactly as they went in — Ideogram 4.5. If your source is a 24-megapixel product photograph and the requirement is that the untouched pixels are literally the untouched pixels, in-place editing at source resolution is the feature you are buying, and Krea 3's 2K-native, 4K-for-print output is not a substitute for it.

• A brand asset that must not drift — Krea 3, with an asterisk. LoRA training on the edit model is the only published mechanism on either side for locking a product or a person, but it is a training step with its own data work, and the model ships in beta.

• Either one, if you need a published score before you commit — wait. Ideogram 4.5's 1,063 Elo is measured and mid-board, with an interval wide enough (1,046 to 1,080) that it overlaps several models around it. Krea 3 has no independent score, and no reviewer has published a matched-settings run of the two against each other.

What would settle this is one matched run: same source images, same prompts, Ideogram 4.5 at High against Krea 3 at 2K, scored by someone who is not selling either one. Until that exists, the price column is real, the resolution column is real, and the quality column has a number on one side and a blank on the other. Buy on what is filled in.