Krea 3 Ultra vs Krea 3: You Are Buying a Mode, Not a Model

The cheapest way to understand the difference between Krea 3 Ultra and Krea 3 is that there is no difference in the model. Both are the same endpoint — POST /generate/image/krea/krea-3 — and the same weights. Ultra is the third value of a parameter called thinking_mode, alongside fast and high. What changed on 20 October 2026 is that Ultra stopped requiring beta access and became callable by any workspace on a Max plan or above; Krea 3 had shipped five days earlier, on 15 October 2026, and came out of beta in the same changelog entry. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

That framing sounds like it makes the comparison trivial — just pick the quality setting. It does not, because Krea's own benchmark pages show Fast beating Ultra on two of the seven editing actions Artificial Analysis measures, and because the price gap between the modes is not the flat multiplier the pricing table suggests. This is a piece about which mode to set, on what job, at what cost.

What each mode actually does

Krea documents the three modes in a single table, and the difference is in the planning, not the rendering:

• fast — expands the prompt and generates. Default, fastest, and Krea's recommendation for "clear, simple edits and most everyday requests".

• high — runs a stronger reasoning pass that plans the edit from the prompt and the reference images before generating. Slower. Intended for multi-step edits, several references to combine, or a fast result that missed the brief.

• ultra — generates several candidate edits, reviews them against your prompt and references, and returns the best one. Always 2K. Slowest; Krea says a job "can take several minutes", with a ceiling of about ten minutes, or sixteen at priority: "low".

Ultra's constraints are the sort that decide whether it is usable before quality enters the conversation. It requires at least one image in image_urls, so text-to-image is out. It will not output below 2K. It is unavailable on workspaces under Max. And on a request carrying Krea's Zero Data Retention header it returns HTTP 422 — an enterprise with a ZDR requirement cannot use Ultra at all, regardless of plan.

High has no such limits. It works at 1K or 2K, takes text-only prompts, and runs under ZDR. If your constraint is data handling rather than quality, the realistic comparison is high against fast, and Ultra never enters it.

The two boards where Fast wins

Artificial Analysis scores Krea 3 in both configurations, which makes this one of the few mode comparisons with independent numbers on both sides. Read on 20 October 2026, Ultra sits at 1150.58 Elo, rank range 3–6, on 5,378 samples. Fast sits at 1125.11, rank range 7–15, on 5,087 samples. Ultra is 25 Elo ahead overall, which is a real gap rather than a tie — the intervals, ±14 each, do not overlap by much.

Then the category boards disaggregate it. Five of seven editing actions go to Ultra, but two do not:

• Live-Action Film — Fast #8 at 1151.6, Ultra #10 at 1135.79. Fast wins by 15.8 Elo.

• Scene & Style Edit — Fast #11 at 1127.35, Ultra #18 at 1119.78. Fast wins by 7.6 Elo.

• Object-Level Edit — Ultra #4 at 1161.08, Fast #5 at 1157.84. Ultra wins by 3.2, inside the noise.

The Live-Action result is the one worth pausing on. Ultra's entire mechanism is to generate several candidates and pick the best against the prompt — and on cinematic film-still edits, that review process produces a worse result than a single fast pass. The most likely explanation is that the selection step is optimising for prompt adherence, and film looks are judged on qualities a prompt describes poorly: grain, falloff, the way light behaves. Whatever the cause, the practical consequence is that "set Ultra and forget it" is wrong for precisely the work Krea's marketing shows most often.

The other side of the ledger is wider than the overall rank implies. On Reasoning-Based Edit, Ultra is #4 and Fast is #20 — a sixteen-place gap on the action that rewards planning. On Marketing & Advertising, Ultra is #4 and Fast is #21. On Animation & Gaming, #2 against #13. Those are the jobs where the candidate-and-review loop is doing what it was built to do.

Category ranks rest on a few hundred votes each — Artificial Analysis's category boards carry between roughly 300 and 800 samples per model — with intervals of ±20 to ±38. Treat a single category placement as a direction and check the interval before you move a production workflow onto it.

What Ultra costs on top of Krea 3

Krea prices by component and the components add. The base is $0.03 per 1K image or $0.06 per 2K. Each 1K reference adds $0.02; each 2K reference adds $0.07. High thinking adds $0.02 per generation. Ultra adds $0.20 per generation, at 2K only.

Three worked cases, all at 2K:

• One hero edit, one 2K reference — Fast $0.06 + $0.07 = $0.13. Ultra $0.13 + $0.20 = $0.33. Ultra costs 2.5× as much.

• A four-reference product composite — Fast $0.06 + $0.28 = $0.34. Ultra $0.54. Ultra costs 1.6× as much.

• A ten-reference campaign set, which exceeds Ultra's practical reference budget of four before Krea starts downscaling every reference to 1536 px on its longest side — the mode is not designed for it, so the comparison does not apply.

The premium shrinks as references grow, because the reference fees are common to both modes and the $0.20 is fixed. A team running one-reference edits pays the full Ultra markup; a team running four pays about 60% more. Neither number is the 3.3× that comparing $0.20 to $0.06 implies, and neither is the "same price" that comparing two $0.06 base prices implies.

Throughput, honestly

Krea's launch claim for Krea 3 is "most edits in under 8 seconds, end to end, at native 2K". That figure belongs to the fast path. The company has not published a median for high, and it describes Ultra only as "several minutes" with a ten-minute ceiling. Anyone budgeting a catalogue run should treat Ultra as a different class of operation: the same 1,000-image job that finishes inside a single working session on fast will not finish on Ultra in one.

That is where a routing layer earns its keep rather than being a line item. OrcaRouter gives one key across a few hundred models with provider list prices passed through at 0% markup, so the arithmetic above is the arithmetic you are billed, and automatic failover when a provider errors or rate-limits is what keeps a long Ultra batch from stalling on a single bad job. Krea 3 itself is not in our catalogue — Ultra is sold through Krea's own API and requires a Krea Max plan to unlock — but the models you would route around it, including O​penAI's gpt-image-2.5 and G​oogle's Nano Banana Pro, are. Which matters for the mode decision: the cheapest test of "does Ultra beat what I have" is to run the same brief through both on one key.

Choosing between them

Set Ultra when the edit requires inference — arithmetic in an image, a scene advanced to a later moment, a receipt that has to recompute its own total — and when one good answer is worth several minutes. Artificial Analysis's published example is exactly this shape: a cart edit that changes a mug from one to three and has to re-derive the line total, the subtotal, the tax and the total. Ultra carried the line item through; two competitors updated the subtotal and left the line stale.

Stay on fast for volume, for anything that has to come back in seconds, and for cinematic or relighting work where the leaderboard says Ultra is not ahead. Reach for high when a fast result missed the brief and the job is too complex for a retry, but you cannot afford Ultra's latency or your workspace cannot meet its plan and data-handling conditions.

The decision worth making explicitly is whether Ultra is a default or an escalation. Krea sells it as a setting, priced per generation, on the same endpoint — which means the rational configuration for most teams is fast as the default and Ultra as a deliberate call on the handful of assets that justify $0.33 and five minutes. The teams that will lose money on this release are the ones that flip the mode globally because it ranks #5, and then find their film-still edits got worse.