Krea 3 Ultra vs Nano Banana 2.1: A Tie on the Board, Two Different Products

One point one Elo. That is the entire measured difference between Krea 3 Ultra at 1150.58 and Nano Banana 2.1 at 1151.65 on Artificial Analysis's image-editing leaderboard, as read on 20 October 2026 — and the two rows carry the same published rank range of 3–6. On the leaderboard, they are the same model with different names on it. 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.

Off the leaderboard they are barely comparable. Krea 3 Ultra is a thinking mode on Krea 3 — an endpoint that returns a job id you poll — which takes several minutes per edit, outputs 2K only, and costs $0.20 on top of the base price every time you use it. Nano Banana 2.1 is G​oogle's high-efficiency workhorse, generally available since 6 October 2026, sold as tokens at $0.0504 per 2K image, capable of 4K, loaded with up to 14 references, and wired into G​oogle Search. One is a scarce, deliberate, expensive call; the other is a metered utility you can hammer.

The useful question is therefore not which scores higher. It is which of those two shapes your job needs — and the price comparison people usually run, $0.33 against $0.05, is measuring the wrong two things. Below, the numbers that actually differ, then a way to decide.

What the tie is made of

Artificial Analysis fits a Bradley-Terry model over pairwise human votes on edited images, on a scale anchored so FLUX.2 [dev] sits at 1000. The 20 October read put Nano Banana 2.1 third at 1151.65 with a ±13 interval on 6,684 samples, and Krea 3 Ultra fifth at 1150.58 with ±14 on 5,378 samples. MAI-Image-2.6 sat between them at 1151.03. Three models, one rank range, intervals overlapping almost completely.

Where the two diverge is by workload, and there the gaps are larger than 1.1 Elo — in both directions:

• Marketing & Advertising — Nano Banana 2.1 1196.4 against Krea 3 Ultra 1182.79. G​oogle ahead by about 14 Elo.

• Retail & Ecommerce — Nano Banana 2.1 1148.92 against 1129.84. G​oogle ahead by 19.

• Animation & Gaming — Krea 3 Ultra #2 at 1178.77; Nano Banana 2.1 below it in the category. Krea ahead at the point estimate.

• Reasoning-Based Edit — Krea 3 Ultra #4 at 1186.82, Nano Banana 2.1 second in the same action at 1193.07. G​oogle ahead by 6.

• Identity Preserving Edit — Nano Banana 2.1 1197.8 against Krea 3 Ultra 1182.47. G​oogle ahead by 15.

• Composition & Framing — Krea 3 Ultra #3 at 1182.68, with Nano Banana 2.1 not in that action's top three.

Read that list as a warning against buying on the headline rank. If your work is advertising creative, product photography, or keeping the same face across a campaign, the board that matters says G​oogle — by margins several times the overall gap. If it is animation, game assets, or edits that need a reasoning step, Krea's case is stronger than #5 suggests.

All category boards carry intervals of ±20 to ±38 on a few hundred votes per model. A 14-Elo category gap is directionally useful and not statistically settled.

Price, done properly

Nano Banana 2.1 is billed in tokens: $1.50 per million input tokens and $30.00 per million image output tokens. G​oogle converts that to about $0.0336 per 1K image, $0.0504 per 2K, and $0.113 per 4K. Those are list prices, and they do not move with the number of references you send.

Krea 3 Ultra is billed by component, and every component stacks: $0.06 base for a 2K image, $0.07 per 2K reference, $0.02 for high thinking if you use it, $0.20 for Ultra. One 2K reference puts a single Ultra edit at $0.33. Four puts it at $0.54. A high-quality 4K Nano Banana 2.1 image with a dozen references is $0.113.

• At one reference, 2K — Krea 3 Ultra $0.33 against Nano Banana 2.1 $0.0504. G​oogle is roughly 6.5× cheaper.

• At four references, 2K — Krea 3 Ultra $0.54 against $0.0504. G​oogle is roughly 10.7× cheaper.

• At 4K — Krea 3 Ultra cannot go there. Nano Banana 2.1 is $0.113.

The comparison that keeps appearing in launch coverage, "$0.06 against $0.135", uses Krea 3's fast mode and Nano Banana Pro's price. Neither side of that sentence is the model this page is about. For Ultra against 2.1, the honest ratio is closer to an order of magnitude.

Latency runs the same direction and by a similar factor. Ultra is a candidate-generating, candidate-reviewing mode: Krea's docs put a job at "several minutes", with a ceiling around ten, and sixteen at low priority. G​oogle publishes no per-image latency for Nano Banana 2.1, but the model is built on the Flash lineage — the previous generation's pitch was speed at volume — and it exposes a thinking-level control with minimal, medium and high settings, defaulting to medium. On throughput, the two products are not in the same market.

Capability differences the Elo cannot see

• Resolution — Krea 3 Ultra is 2K only. Nano Banana 2.1 does 1K, 2K and 4K, with a published per-aspect-ratio pixel table. G​oogle's Flash sibling, Nano Banana 2, adds a 512px tier that 2.1 does not have.

• References — Nano Banana 2.1 accepts up to 14 reference images, holding character resemblance for up to four characters and fidelity for up to ten objects in one workflow, per G​oogle's docs. Krea 3 accepts up to eight in general, but with five or more each reference is downscaled to 1536 px on its longest side; Krea advises sending four or fewer for full detail. Ultra requires at least one reference.

• Search grounding — Nano Banana 2.1 can ground generation in G​oogle Web Search and, new in this generation, G​oogle Image Search, so a prompt can reference a real current product or a specific retrieved image. Krea 3 has nothing equivalent.

• Video input — Nano Banana 2.1 accepts a video as context and produces a still from it. Krea 3 takes images and text.

• Watermarking — every image G​oogle generates carries a SynthID watermark. Krea's docs do not describe an equivalent for Krea 3.

• Data handling — Krea 3 supports Zero Data Retention for enterprise workspaces, but Ultra returns HTTP 422 under a ZDR header, so enterprise ZDR and Ultra are mutually exclusive. G​oogle's handling is governed by its own API terms.

• Regional edits — Krea 3 offers regional_edit: true, which regenerates only the changed region and leaves everything else pixel-exact, so you can chain edits on the latest output without artefact build-up. That is a genuine product difference, and the kind of thing a preference leaderboard does not capture.

Choosing, on a real brief

The two models are strongest on different jobs, and the leaderboard tie makes that easy to miss.

Take a campaign that needs one hero image per market, with the same product label and the same model in every frame, at print resolution, referencing a dozen source images. Nano Banana 2.1 wins on every axis that matters: 4K, 14 references, identity preservation ranked 15 Elo above Krea, $0.11 an image, and throughput that makes a hundred variations an afternoon's work. Krea 3 Ultra would cost $0.54 a frame, take minutes each, and downscale most of your references.

Take a batch of product packshots that must keep their label text and logo pixel-exact while the scene around them changes, edited iteratively from the previous output. Krea's regional edit mode is built for exactly that, its Retail & Ecommerce rank (#4 in the use case, behind Nano Banana 2.1 but ahead of most of the field) is respectable, and 2K is usually enough. Ultra is the wrong mode here for a different reason: Fast at $0.13 with one reference does the job, and Fast is #11 overall — still ahead of Nano Banana Pro's 1106.

Nano Banana 2.1 is the default for most teams because it is cheap, fast, high-resolution, and references more inputs than any competitor. Krea 3 Ultra is a specialist call: a few minutes and roughly 30 cents for one edit that has to be right, on a job where the model's own candidate review does better than a single pass. The board says they are equally good. The pricing says one of them is a utility and one is a consultant.

One key for the comparison

The reason this matchup is hard to settle from outside is that the two models are billed in incompatible units — per generation with surcharges on one side, per token with a resolution ladder on the other — measured across two different latency classes. The only reliable way to compare is to run the same ten briefs through both and count what the invoice says.

OrcaRouter is built for that: one API over a few hundred models at 0% markup, with provider list prices passed through, automatic failover when a provider errors or rate-limits, and a routing DSL for sending different jobs to different models inside one workflow. Nano Banana 2.1 is G​oogle's own endpoint and is not in our catalogue today; what we do route includes Nano Banana 2 (google/gemini-3.1-flash-image-preview), Nano Banana Pro, and O​penAI's gpt-image-2.5 — which makes the useful first test a three-way one: does Ultra's $0.33 edit beat a $0.05–0.15 G​oogle or O​penAI edit on your actual brief? Krea 3 stays on Krea's API either way, because Ultra is sold there.

Until someone runs that test, the defensible summary is narrow and unglamorous: on 20 October 2026 these two models scored the same, and on everything else that decides a production choice — price, resolution, references, latency, data handling — they are not close. Pick the shape of the job first. The rank will not help you.