Three independent leaderboards published a Krea 3 benchmark on October 19 and 20, 2026. They do not agree on a single rank, and they should not be collapsed into one.
As of those snapshots:
- Arena, Multi-Image Edit: Krea 3 Ultra is #4, at 1435±11, marked Preliminary, on 2,991 votes. Rank spread 4–5. The page is dated October 19, 2026.
- Arena, Image Edit (one input image): Krea 3 Ultra is #8, at 1411±11, marked Preliminary, on 2,619 votes. Rank spread 5–10. Same page date.
- Artificial Analysis, AA-Image-Editing v2.1: Krea 3 Ultra is #5 at 1151 Elo (1150.58 exact), rank range 3–6, statistically tied for #3 with Nano Banana 2.1 and MAI-Image-2.6. Krea 3 Fast is #11 at 1125 Elo (1125.11 exact), ahead of Nano Banana 2 and Nano Banana Pro on the point estimate. Live page read about 10:45 AM PT on October 20, 2026.
- Design Arena, Image Editing: Krea 3 (Ultra Thinking) ranks #4 of 51 at 1298 Elo, which makes Krea the #2 image-editing lab after OpenAI. It edges out Reve 2.0, and ranks #2 in Architecture and #3 in Landscape. Thread.
The chart under the title is Arena’s Multi-Image Edit graphic. For the model itself, see Introducing Krea 3.
How to read these benchmarks
An AI image editing leaderboard, on all three of these sites, is a blind vote. A person sees two edits made from the same input and the same instruction, and picks the one they prefer. The model names stay hidden until the vote is cast.
Those votes become a rating. Arena and Artificial Analysis fit a Bradley-Terry model and report it as Elo. Design Arena does the same, and converts strength to a rating with 400 × log₁₀(strength). A higher score means the model wins more of the matchups it is in. The number only means something on that board. Arena’s 1435, Artificial Analysis’s 1151, and Design Arena’s 1298 are not three readings of one scale. Artificial Analysis anchors its scale so FLUX.2 [dev] sits at 1000. Design Arena’s API docs describe a base of 1200. Subtracting one board’s Elo from another’s does not produce a gap.
The ± figure is a confidence interval. Arena’s ranking notes turn the intervals into a rank spread: the best and worst rank a model can hold before another model’s interval clears it. Overlapping spreads are a tie. A point rank of #4 with a spread of 4–5 means the order against the neighbor is not settled. Artificial Analysis publishes a rank range on the same idea. Design Arena’s chart does not print intervals or vote counts.
Sample size is the other half. Arena’s Preliminary tag means the model is early in its vote count. Krea 3 Ultra has 2,991 votes on Multi-Image Edit and 2,619 on Image Edit, against tens or hundreds of thousands for models that have been listed longer. Several category pages sit on a few hundred votes, with intervals of ±20 to ±38. Read a category rank as a direction, and read the interval before you treat the gap as settled.
Three leaderboards, side by side
Scores in this table are not comparable across rows. Each row is dated on its own.
| Leaderboard | Board | Krea 3 result | Score | Sample | As of |
|---|---|---|---|---|---|
| Arena | Multi-Image Edit | #4, spread 4–5, Preliminary | 1435±11 | 2,991 votes | Page dated October 19, 2026. 45 models, 9,408,830 votes on the board |
| Arena | Image Edit (single image) | #8, spread 5–10, Preliminary | 1411±11 | 2,619 votes | Page dated October 19, 2026. 60 models, 31,404,772 votes on the board |
| Artificial Analysis | AA-Image-Editing v2.1 | Krea 3 Ultra #5, range 3–6 | 1151 Elo (1150.58), ±14, CI 1137–1165 | 5,378 samples | Live page, about 10:45 AM PT, October 20, 2026. 74 models |
| Artificial Analysis | AA-Image-Editing v2.1 | Krea 3 Fast #11, range 7–15 | 1125 Elo (1125.11), ±14, CI 1111–1139 | 5,087 samples | Same read |
| Design Arena | Image Editing | #4 of 51. #2 image-editing lab after OpenAI. #2 Architecture, #3 Landscape | Elo 1298 | Votes not printed on the chart | October 20, 2026 |
Arena also said, in the same post, that Krea is #4 by lab on Multi-Image Edit and #5 by lab on Image Edit. Design Arena ranks Krea as the #2 image-editing lab after OpenAI. Artificial Analysis calls Krea 3 Ultra “the highest-ranked image editing model from outside OpenAI, Google and Microsoft.” Those are lab claims from the posts. The model ranks in the table are the page ranks.
Arena
Arena (arena.ai) added Krea 3 Ultra to the Single-Image Edit arena and the Multi-Image Edit arena on October 20, 2026. The changelog for that day says so. The leaderboard pages themselves are stamped October 19, 2026. The post is here, and the follow-up links the multi-image board.
In battle mode, a voter sees two anonymous models and picks a preferred edit. The names are revealed after the vote. Multi-Image Edit has been its own category since January 23, 2026: those are the votes where more than one image was passed in. As of July 30, 2026, both edit arenas use the same seven categories as text-to-image.
"Krea 3 Ultra by @krea_ai has landed in the Single-Image Edit Arena and the Multi-Image Edit Arena. @krea_ai is now among the top 5 labs for both arenas: #4 by lab in Multi-Image Edit, and #5 by lab in Image Edit!"
Multi-Image Edit: #4
Page header, as of October 19, 2026: 9,408,830 votes, 45 models. Board.
| Rank | Spread | Model | Score | Votes |
|---|---|---|---|---|
| 1 | 1–1 | GPT-Image-2.5 Sunburst (OpenAI) | 1518±6 Preliminary | 39,653 |
| 2 | 2–2 | GPT-Image-2.5 Flare (OpenAI) | 1479±6 Preliminary | 39,664 |
| 3 | 3–3 | GPT-Image-2 (Medium) (OpenAI) | 1453±4 | 142,868 |
| 4 | 4–5 | Krea 3 Ultra (Krea) | 1435±11 Preliminary | 2,991 |
| 5 | 4–5 | Gemini Nano Banana 2.1 (Google) | 1432±6 Preliminary | 10,284 |
| 6 | 6–6 | Seedream 5.0 Pro (Bytedance) | 1412±4 | 88,171 |
| 7 | 7–7 | Muse Image (Meta) | 1399±5 | 108,047 |
| 8 | 8–10 | Gemini 3.1 Flash Image (nano-banana-2, Web Search) | 1367±4 | 112,463 |
| 9 | 8–10 | Gemini 3 Pro Image Preview (nano-banana-pro) | 1365±4 | 298,134 |
| 10 | 8–10 | Gemini 3 Pro Image (nano-banana-pro, 2K) | 1363±3 | 272,470 |
| 11 | 11–12 | ChatGPT Image Latest (High Fidelity, 20251216) | 1351±3 | 255,541 |
| 12 | 11–14 | Reve 2.0 | 1344±7 | 10,323 |
| 13 | 12–14 | GPT-Image-1.5 (High Fidelity) | 1342±3 | 238,717 |
| 14 | 12–14 | Ideogram 4.5 | 1336±7 Preliminary | 6,722 |
| 15 | 15–15 | UNI-1.1 Max (Luma AI) | 1308±5 | 21,994 |
Krea 3 Ultra sits 18 points below GPT-Image-2 (Medium) and 3 points above Nano Banana 2.1. The rank spread of 4–5 overlaps Nano Banana 2.1’s spread of 4–5, so Arena’s own rule treats the two as tied for #4. Krea has the fewest votes in this top 15, which is why the row is marked Preliminary.
Multi-Image Edit, top 15
Arena's graphic, posted October 20, 2026. Scores match the Multi-Image Edit page dated October 19, 2026. Krea 3 Ultra is the highlighted bar at 1435.
Multi-Image Edit, by category
All four category pages below are dated October 19, 2026. Each board has 33 models. The ranks match the sentence in Arena’s post. The intervals are wide because the vote counts are small.
| Category | Krea 3 Ultra | Spread | Votes | Above | Below | Board votes |
|---|---|---|---|---|---|---|
| Commercial Design | #2, 1534±31 Preliminary | 1–4 | 455 | #1 GPT-Image-2.5 Sunburst 1558±10 | #3 GPT-Image-2.5 Flare 1507±9 | 327,359 |
| 3D Imaging & Modeling | #2, 1486±33 Preliminary | 2–5 | 346 | #1 Sunburst 1531±11 | #3 Flare 1482±11 | 178,826 |
| Cartoon / Anime / Fantasy | #4, 1473±20 Preliminary | 3–5 | 774 | #3 GPT-Image-2 (Medium) 1480±6 | #5 Nano Banana 2.1 1452±12 | 388,629 |
| Text Rendering | #4, 1507±36 Preliminary | 2–5 | 302 | #3 GPT-Image-2 (Medium) 1509±7 | #5 Nano Banana 2.1 1502±15 | 253,264 |
Commercial Design is the sharpest of these. The point score, 1534, is above Flare and inside the interval of Sunburst: the spread runs from 1 to 4, on 455 votes. 3D is #2 on the point rank and overlaps Flare (1486±33 against 1482±11), on 346 votes. Text Rendering’s 1507±36 sits 2 points under GPT-Image-2 (Medium) and 5 points over Nano Banana 2.1, on 302 votes. Cartoon is the best-sampled of the four, at 774 votes, and the spread is 3–5.
Single-image edit: #8
Arena’s Image Edit board is the single-image arena. The main page and the single-image URL returned the same table. Header, as of October 19, 2026: 31,404,772 votes, 60 models.
| Rank | Spread | Model | Score | Votes |
|---|---|---|---|---|
| 1 | 1–1 | GPT-Image-2.5 Sunburst | 1514±6 Preliminary | 82,078 |
| 2 | 2–2 | GPT-Image-2.5 Flare | 1481±5 Preliminary | 79,506 |
| 3 | 3–3 | GPT-Image-2 (Medium) | 1461±3 | 328,307 |
| 4 | 4–4 | Grok Imagine Image 2.0 (SpaceXAI) | 1439±6 Preliminary | 30,542 |
| 5 | 5–7 | Gemini Nano Banana 2.1 | 1426±6 Preliminary | 12,938 |
| 6 | 5–7 | MAI-Image-2.6 (Microsoft AI) | 1425±5 | 69,611 |
| 7 | 7–8 | Grok Imagine Image 2.0 (20260801) | 1415±5 | 76,730 |
| 8 | 5–10 | Krea 3 Ultra | 1411±11 Preliminary | 2,619 |
| 9 | 8–10 | MAI-Image-2.5 | 1400±4 | 190,119 |
| 10 | 8–12 | Muse Image (Meta) | 1398±4 | 164,982 |
| 11 | 10–16 | Seedream 5.0 Pro | 1391±4 | 307,763 |
| 12 | 10–16 | Grok Imagine Image (Quality, 20260519) | 1391±6 | 38,207 |
| 13 | 11–16 | Gemini 3 Pro Image (nano-banana-pro, 2K) | 1390±3 | 676,298 |
| 14 | 11–16 | Gemini 3.1 Flash Image (nano-banana-2, Web Search) | 1387±3 | 253,535 |
| 15 | 11–16 | ChatGPT Image Latest (High Fidelity, 20251216) | 1387±3 | 651,019 |
The point rank is #8. The spread of 5–10 is the number to read with it: on 2,619 votes and a ±11 interval, Krea 3 Ultra still overlaps the cluster from Nano Banana 2.1 and MAI-Image-2.6 down through Muse Image. It is 4 points under Grok Imagine Image 2.0 (20260801) and 11 points over MAI-Image-2.5. It is not a top-four result on this board. Grok Imagine Image 2.0 (SpaceXAI), Nano Banana 2.1, and MAI-Image-2.6 all sit above the point rank.
Image Edit, top 15
Arena's single-image graphic, posted October 20, 2026. A few labels are shorter than the page: rank 4 is shown as Grok Imagine Image 2.0 (canvas), rank 7 as Grok Imagine Image 2.0 (low), and ranks 13 and 14 as Nano Banana Pro 2K (1390) and Nano Banana 2 (1387). The scores match the page dated October 19, 2026.
Single-image edit, by category
Dated October 19, 2026. Each board has 45 models. Arena’s post lists the same four ranks: #6 in Commercial Design, Cartoon/Fantasy, and Text Rendering, and #8 in 3D.
| Category | Krea 3 Ultra | Spread | Votes | Above | Below | Board votes |
|---|---|---|---|---|---|---|
| Commercial Design | #6, 1448±31 Preliminary | 4–13 | 361 | #5 Grok Imagine Image 2.0 1469±10 | #7 MAI-Image-2.6 1441±8 | 875,534 |
| Cartoon / Anime / Fantasy | #6, 1452±22 Preliminary | 4–8 | 679 | #5 Grok Imagine Image 2.0 1453±8 | #7 Grok Imagine Image 2.0 (20260801) 1441±7 | 1,093,251 |
| Text Rendering | #6, 1481±38 Preliminary | 4–10 | 229 | #5 Grok Imagine Image 2.0 1490±11 | #7 Reve 2.1 1454±13 | 651,976 |
| 3D Imaging & Modeling | #8, 1437±33 Preliminary | 4–14 | 298 | #7 Grok Imagine Image 2.0 (20260801) 1443±9 | #9 MAI-Image-2.5 1427±7 | 531,909 |
Cartoon is one point under the model at #5 (1452±22 against 1453±8), on 679 votes. Text Rendering’s interval is ±38, on 229 votes, so the #6 is a point estimate with a spread of 4–10. Commercial Design’s spread runs from 4 to 13.
What the Arena results mean for a creator
Multi-Image Edit is the closer match for a job that starts from several pictures: a product, a person, a material, a scene. That is where Krea 3 Ultra’s point rank is #4 overall, and #2 in Commercial Design and in 3D. Those category #2s are real rows on the pages, and they rest on 455 and 346 votes. Treat them as early, not as a finished gap.
Single-image edit is a deeper field, with 60 models and a much larger vote pool. #8, with a spread of 5–10, means the model is in the top group and not separated from it. If the work is one photo and one instruction, this is the Arena number to quote. If the work is a composite from references, quote the multi-image number.
The lab ranks, #4 and #5, come from Arena’s post. This post does not re-derive them from the labs view.
Artificial Analysis
Artificial Analysis posted its Krea 3 thread on October 20, 2026, at 9:13 AM PT. The opening post puts Krea 3 Ultra at #5 on AA-Image-Editing v2.1. The last post links the editing leaderboard. The charts in the thread are stamped “ELO AS OF 2026-10-16”. The category table later in this section is the live page, read about 10:45 AM PT on October 20, not the October 16 chart.
"Krea 3 Ultra lands at #5 on AA-Image-Editing v2.1, with quality similar to Nano Banana 2.1 and MAI-Image-2.6."
What AA-Image-Editing measures
The leaderboard ranks models on editing an image from a text instruction, by Elo from pairwise human preference. Voters compare two edits of the same input and the same instruction and pick the one they prefer. Artificial Analysis describes the judges as validated private evaluators. Identities are hidden. There is a minimum viewing time before a vote counts.
All votes are fit at once with a Bradley-Terry model and reported as Elo. The anchor is FLUX.2 [dev] at 1000. Confidence intervals treat each vote as independent, so Artificial Analysis notes that they may be slightly narrow. Prompts are sampled evenly across a taxonomy of 10 use cases and 7 editing actions, and the prompt set is refreshed every month. The overall board in this read has 74 models. Category boards have 71 to 74.
Artificial Analysis said it tested both Krea modes through the Krea API at 2K resolution. Krea 3 (Ultra) is recorded as released October 15, 2026, and added to the board at 01:21 UTC that day (6:21 PM PT on October 14). The page lists both Krea 3 (Fast) and Krea 3 (Ultra) among 11 models added in the prior 30 days.
Overall: Ultra #5, Fast #11
Live page, about 10:45 AM PT on October 20, 2026. The page shows no snapshot date of its own. Exact Elo is the value in the page data. The first number is what the page displays.
| Rank | Range | Model | Elo | 95% CI | Samples |
|---|---|---|---|---|---|
| 1 | 1 | GPT Image 2.5 Sunburst (max) | 1199 (1198.71) | ±11, 1188–1210 | 18,059 |
| 2 | 2 | GPT Image 2.5 Flare (max) | 1181 | ±11, 1170–1192 | 18,954 |
| 3 | 3–6 | Nano Banana 2.1 | 1152 (1151.65) | ±13, 1139–1165 | 6,684 |
| 4 | 3–6 | MAI-Image-2.6 | 1151 (1151.03) | ±12, 1139–1163 | 16,468 |
| 5 | 3–6 | Krea 3 (Ultra) | 1151 (1150.58) | ±14, 1137–1165 | 5,378 |
| 6 | 3–8 | MAI-Image-2.6-Flash | 1143 (1143.12) | ±13, 1130–1156 | 10,579 |
| 7 | 6–12 | GPT Image 2 (high) | 1136 | ±13, 1123–1149 | 14,367 |
| 8 | 6–13 | Seedream 5.0 Pro | 1132 | ±13, 1119–1145 | 13,183 |
| 9 | 7–14 | Muse Image | 1129 | ±12, 1117–1141 | 13,159 |
| 10 | 7–15 | MAI-Image-2.5 | 1128 (1128.26) | ±13, 1115–1141 | 13,347 |
| 11 | 7–15 | Krea 3 (Fast) | 1125 (1125.11) | ±14, 1111–1139 | 5,087 |
| 12 | 8–15 | Nano Banana 2 (Gemini 3.1 Flash Image) | 1124 (1123.64) | ±12, 1112–1136 | 16,454 |
| 13 | 8–15 | MAI-Image-2.5-Pro | 1122 | ±12, 1110–1134 | 15,452 |
| 14 | 9–16 | GPT Image 1.5 (high) | 1119 | ±13, 1106–1132 | 13,319 |
| 15 | 9–16 | Grok Imagine Image 2.0 | 1117 | ±13, 1104–1130 | 8,490 |
| 16 | 15–17 | Nano Banana Pro (Gemini 3 Pro Image) | 1106 | ±12, 1094–1118 | 15,358 |
Krea 3 Ultra’s displayed Elo is 1151. So is MAI-Image-2.6. Nano Banana 2.1 displays as 1152. The exact order is Nano Banana 2.1 at 1151.65, MAI-Image-2.6 at 1151.03, and Krea 3 Ultra at 1150.58. All three publish a rank range of 3–6, and the intervals overlap (1137–1165, 1139–1165, and 1139–1163). That is a statistical tie for #3. It matches Artificial Analysis’s own line, “quality similar to Nano Banana 2.1 and MAI-Image-2.6.” The displayed rank is still #5. MAI-Image-2.6-Flash, at 1143.12, has a wider range of 3–8 and a lower point score.
A card in Krea’s launch thread shows Ultra at #4 on this board, with MAI-Image-2.6 at #5. The leaderboard page and Artificial Analysis’s own chart both put MAI-Image-2.6 at #4 and Krea 3 Ultra at #5. This post uses the leaderboard.
Krea 3 Fast, the default mode, is #11 at 1125.11. Nano Banana 2 is #12 at 1123.64. Nano Banana Pro is #16 at 1106. Fast’s point estimate is ahead of both. The intervals overlap: Fast is 1111–1139, Nano Banana 2 is 1112–1136, and Nano Banana Pro is 1094–1118. The gap to Nano Banana 2 is inside the noise. The gap to Nano Banana Pro is larger, and the intervals still touch.
AA-Image-Editing v2.1, top 15
Final frame of Artificial Analysis's animated chart, posted October 20, 2026. The names are readable: Krea 3 Ultra at #5 and Krea 3 Fast at #11. The bar values fade, so the Elo figures in the table come from the leaderboard page, not from this frame.
Cost
On krea.ai/krea-3, Krea 3 is 10× faster, 3× cheaper, compared to GPT Image 2 and Nano Banana Pro.
All 17 categories
Artificial Analysis splits image editing into 10 use cases and 7 editing actions. Rank below is the displayed rank on the live page. Elo is exact. “Gap” is Ultra’s Elo minus the category leader, which is what the radar charts label, rounded. “n” is Ultra’s sample count. Each category board has 71 to 74 models.
As of the October 20, 2026 read, Krea 3 Ultra is in the top 7 in 14 of 17 categories, and in the top 7 on 5 of 7 editing actions. The three outside the top 7 are Live-Action Film (#10), Enhancement & Restoration (#8), and Scene & Style Edit (#18).
The thread posted that morning says something slightly different, and it should be quoted with its date. Artificial Analysis wrote that Ultra “places in the top 7 in 15 of our 17 Image Editing categories” and “ranks in the top 7 on 6 of the 7 editing actions.” The charts attached to that thread are stamped Elo as of October 16, 2026. The Elo gaps on those charts still match the gaps below. A few ranks moved between the chart and the later page read. Where they differ, the table uses the October 20 page.
Use cases
Use cases are the workloads. Animation & Gaming is game assets, concept art, character design, and comics. Frontier is work at and beyond the current frontier. Marketing & Advertising is ad creatives, campaign visuals, posters, and brand imagery. The other names are the jobs they sound like.
| Category | Ultra | Ultra Elo | 95% CI | n | Gap to #1 | Fast | Fast Elo | Leader | Around Ultra |
|---|---|---|---|---|---|---|---|---|---|
| Animation & Gaming | #2 | 1178.77 | 1135.77–1221.77 | 552 | −12.6 | #13 | 1115.73 | Sunburst 1191.37 | Above: Sunburst. Below: Flare 1166.97 |
| Frontier | #2 | 1169.73 | 1126.73–1212.73 | 546 | −28.5 | #5 | 1150.74 | Sunburst 1198.27 | Above: Sunburst. Below: MAI-Image-2.6-Flash 1155.29 |
| Productivity & Knowledge Work | #4 | 1200.98 | 1156.98–1244.98 | 543 | −24.9 | #14 | 1159.6 | Sunburst 1225.89 | Above: Flare 1220.09. Below: Muse Image 1196.07 |
| Retail & Ecommerce | #4 | 1129.84 | 1086.84–1172.84 | 487 | −52.3 | #12 | 1100.27 | Sunburst 1182.1 | Above: Nano Banana 2.1 1148.92. Below: Seedream 5.0 Pro 1127.32 |
| Social Media & Creator Content | #4 | 1156.24 | 1112.24–1200.24 | 495 | −57.1 | #6 | 1152.75 | Sunburst 1213.36 | Above: MAI-Image-2.6 1156.91. Below: Nano Banana 2.1 1155.07 |
| Marketing & Advertising | #4 | 1182.79 | 1139.79–1225.79 | 530 | −58.6 | #21 | 1112.45 | Sunburst 1241.43 | Above: Nano Banana 2.1 1196.4. Below: MAI-Image-2.6 1179.67 |
| Architecture & Real Estate | #7 | 1083.11 | 1042.11–1124.11 | 540 | −63.2 | #13 | 1069 | Sunburst 1146.32 | Above: GPT Image 2 (high) 1088.5. Below: GPT Image 1.5 (high) 1082.9 |
| Consumer | #7 | 1157.36 | 1115.36–1199.36 | 550 | −65.2 | #10 | 1144.09 | Flare (max) 1222.52 | Above: MAI-Image-2.5 1158.34. Below: MAI-Image-2.6-Flash 1156.01 |
| UIUX Design | #7 | 1136.74 | 1094.74–1178.74 | 551 | −71.6 | #9 | 1125.18 | Sunburst 1208.38 | Above: Nano Banana 2.1 1142.81. Below: MAI-Image-2.5 1135.48 |
| Live-Action Film | #10 | 1135.79 | 1092.79–1178.79 | 538 | −74.3 | #8 | 1151.6 | Flare (max) 1210.1 | Above: Muse Image 1150.74. Below: GPT Image 1.5 (high) 1135.18 |
The thread’s best-result line holds on this later read: Ultra is #2 in Animation & Gaming and #2 in Frontier, behind GPT Image 2.5 Sunburst (max) in both. It is #4 in Productivity & Knowledge Work, Retail & Ecommerce, Social Media & Creator Content, and Marketing & Advertising.
Two Fast results are easy to miss. On Live-Action Film, Fast at #8 (1151.6) ranks above Ultra at #10 (1135.79). On Frontier, Fast is #5, close to Ultra’s #2. The largest Ultra-over-Fast gaps on this page are Marketing & Advertising (#4 vs #21) and Animation & Gaming (#2 vs #13). The thread had cited Marketing as #4 vs #20 for Fast. The October 20 page shows Fast at #21.
Social Media is a tight cluster around Ultra: MAI-Image-2.6 at 1156.91, Ultra at 1156.24, Nano Banana 2.1 at 1155.07. Consumer is the same shape: MAI-Image-2.5 at 1158.34, Ultra at 1157.36, MAI-Image-2.6-Flash at 1156.01. A one-point rank in that band is not a separation.
Gap to the use-case frontier
Artificial Analysis radar, stamped Elo as of October 16, 2026. Labels are the rounded gap from Krea 3 Ultra to the category leader: Animation and Gaming −13, Frontier −29, Productivity −25, Retail −52, Social −57, Marketing −59, Architecture −63, Consumer −65, UIUX −72, Live-Action −74. The October 20 Elo gaps match these labels. Fast is the dotted series.
Editing actions
Editing actions are what the instruction asks the model to do. Composition & Framing is cropping, extending, reframing, and changes to camera or perspective. Reasoning-Based Edit is an instruction that needs an inference about the scene, not a literal swap. Scene & Style Edit is relighting, restyling, and changes to the background, setting, or overall treatment.
| Category | Ultra | Ultra Elo | 95% CI | n | Gap to #1 | Fast | Fast Elo | Leader | Around Ultra |
|---|---|---|---|---|---|---|---|---|---|
| Composition & Framing | #3 | 1182.68 | 1145.68–1219.68 | 727 | −67.3 | #5 | 1164.08 | Sunburst 1249.97 | Above: Flare 1213.46. Below: GPT Image 2 (high) 1177.17 |
| Reasoning-Based Edit | #4 | 1186.82 | 1150.82–1222.82 | 786 | −19.3 | #20 | 1108.11 | Sunburst 1206.16 | Above: Nano Banana 2.1 1193.07. Below: GPT Image 2 (high) 1177.62 |
| Object-Level Edit | #4 | 1161.08 | 1126.08–1196.08 | 811 | −55.5 | #5 | 1157.84 | Sunburst 1216.57 | Above: Flare 1178.18. Below: Krea 3 (Fast) 1157.84 |
| Identity Preserving Edit | #4 | 1182.47 | 1146.47–1218.47 | 776 | −49.0 | #11 | 1159.09 | Sunburst 1231.44 | Above: Nano Banana 2.1 1197.8. Below: GPT Image 2 (high) 1181.25 |
| Text or Symbol Edits | #6 | 1151.33 | 1114.33–1188.33 | 758 | −43.8 | #16 | 1105.01 | Sunburst 1195.1 | Above: MAI-Image-2.6 1156.88. Below: Nano Banana 2 1144.31 |
| Enhancement & Restoration | #8 | 1094.72 | 1059.72–1129.72 | 767 | −54.5 | #12 | 1075.39 | Sunburst 1149.22 | Above: MAI-Image-2.6-Flash 1098.27. Below: MAI-Image-2.5-Flash 1094.04 |
| Scene & Style Edit | #18 | 1119.78 | 1084.78–1154.78 | 753 | −64.4 | #11 | 1127.35 | Sunburst 1184.2 | Above: MAI-Image-2.5-Flash 1122.7. Below: Nano Banana 2 1117.5 |
Five of the seven actions are inside the top 7. Composition & Framing is the best action rank, at #3, and Fast is #5 on the same action. Reasoning-Based Edit is #4 for Ultra and #20 for Fast, the widest mode gap in the action set, with Ultra 19.3 Elo off the leader. Object-Level Edit has Fast directly behind Ultra, #5 at 1157.84 against #4 at 1161.08.
The two actions outside the top 7 are the ones the October 16 chart and the October 20 page disagree on by a place. Enhancement & Restoration is #8. The thread’s “6 of 7 editing actions in the top 7” does not fit a #8. Scene & Style Edit is #18. The thread said #17. Fast ranks ahead of Ultra on Scene & Style, #11 at 1127.35 against #18 at 1119.78.
Gap to the editing-action frontier
Artificial Analysis radar, stamped Elo as of October 16, 2026. Rounded gaps from Krea 3 Ultra to the leader: Reasoning −19, Text −44, Identity −49, Object −55, Enhancement −55, Scene and Style −64, Composition −67. The October 20 gaps match. Fast is dotted. Composition is the best rank and still a large Elo gap, because Sunburst leads that action by a wide margin.
Example edits Artificial Analysis published
The comparisons below are from the thread. Prompts are quoted as they appear on the images. Artificial Analysis did not write a winner’s note on the shopping-cart edit. The description of that one is what the published panels show.
The shopping cart that has to update its own total
Reasoning-Based Edit, UIUX Design, flat design. Prompt on the image: Change the quantity of the mug from 1 to 3, and rename it from Mug to Handmade ceramic espresso mug, set of two. Update every number that depends on the change: the mug line total, the subtotal, the 10% tax, and the total. In the published comparison, the Krea 3 Ultra panel updates the mug line, the subtotal, the tax, and the total for a quantity of three. The GPT Image 2.5 Sunburst (max) and Nano Banana 2.1 panels leave the original mug line in place while still showing the updated subtotal.
The updated mug line is three times the line the other panels still show, which is the quantity change the prompt asks for, from 1 to 3. Ultra then adds the 10% tax on top of the new subtotal. On this one image, Ultra is the panel that carries the line total through. It is one example from the Reasoning-Based Edit set, where Ultra’s category rank is #4 and Fast’s is #20. It is not a claim about every receipt.
Fifteen minutes into the meal
Reasoning-Based Edit, Consumer, photorealistic. The prompt asks to show the grill about fifteen minutes into the meal: pork on the right cooked through, browned, and a bit smaller, rendered fat on the stone, a few pieces gone, garlic golden, kimchi darker where it meets the stone, and the center sauce lower. The image truncates the prompt.
The after slide of a cleanup day
Reasoning-Based Edit, Social Media and Creator Content, photorealistic. Prompt: Advance this frame to the end of the cleanup day so it works as the after slide of the recap post. Remove every can, wrapper and scrap of litter from the grass, swap the loose bags for knotted, visibly full trash bags piled at the group's feet, add grime and grass stains to the gloves and shirt fronts, and shift the light to late afternoon with long shadows stretching from a low sun.
An energy drink, endorsed by gods
Reasoning-Based Edit, Marketing and Advertising, digital painting. Prompt: These figures are endorsing a bright green energy drink that will give you superpowers. Krea 3 Ultra's panel adds the headline UNLEASH GOD-TIER SUPERPOWERS. Ultra's Marketing rank on the October 20 page is #4. Fast's is #21.
A coin pile, brought back into focus
Enhancement and Restoration, Frontier, macro photography. Prompt: Bring the whole pile of coins into crisp high-resolution detail, recovering the engraving so legends like ONE PENNY and ELIZABETH II D G read sharply all the way out to the soft corners of the frame. Clean the fine grain out of the dark gaps between the coins without smoothing away the milled edges, hairline scratches and worn relief on the metal. Enhancement and Restoration is Ultra's #8 action.
A shrine, three hundred years later
Reasoning-Based Edit, Animation and Gaming, 3D render. Prompt: Show this shrine asset the way it would look three hundred years after the last person walked its ramps, left out in the open weather the whole time. The red inlays should have long since gone dark and cold. Animation and Gaming is Ultra's #2 use case.
What the Artificial Analysis results mean for a creator
Ultra is the mode that shows up when the edit has to infer. Reasoning-Based Edit is #4, 19.3 Elo off the leader, and it is the action where Fast falls to #20. The cart, the meal, the cleanup, and the shrine are that kind of instruction: the picture has to change in ways the prompt only implies. Composition & Framing, at #3, is the other action to remember. Crops, extends, and camera moves are where Ultra’s rank is strongest, even though the Elo gap to Sunburst on that action is large (−67.3) because the leader is far ahead of the field.
The weak actions are specific. Scene & Style Edit, which covers relighting and restyling, is #18 for Ultra, and Fast is closer there at #11. Enhancement & Restoration is #8. Live-Action Film is #10 for Ultra and #8 for Fast. If the job is a cinematic relight of a photograph, this board does not put Ultra at the front. If the job is concept art, a character, a poster, a product composite, or an edit that has to stay consistent with a count or a label, Ultra’s category ranks are the relevant ones: #2 in Animation & Gaming and Frontier, #4 in Marketing, Retail, Social, and Productivity.
Fast is not a lesser copy of the same score. It is #11 overall, in a rank range of 7–15, with a point Elo above Nano Banana 2 and Nano Banana Pro and an interval that overlaps both. It stays near Ultra on object-level edits and on composition, and it beats Ultra on live-action and on scene and style. The mode gap is largest where the instruction stops being literal.
Design Arena
Krea 3 (Ultra Thinking) ranks #4 of 51 on Design Arena’s Image Editing leaderboard, at 1298 Elo. That makes Krea the #2 image-editing lab after OpenAI. It edges out Reve 2.0 (by Reve). It ranks even higher in two subcategories: #2 in Architecture and #3 in Landscape.
Design Arena posted that result on October 20, 2026. The thread is here: Design Arena on X.
"Krea 3 (Ultra Thinking) by @krea_ai lands at #4 of 51 on Design Arena's Image Editing leaderboard, making Krea the #2 Image Editing lab after @OpenAI. Krea 3 (Ultra Thinking) edges out Reve 2.0 by @reve and climbs even higher in key subcategories, ranking #2 in Architecture and #3 in Landscape."
The leaderboard
Design Arena ranks models with blind tournaments. Each session shows four models, plus a backup, all given the same prompt, and turns that session into five pairwise battles. Names stay hidden until the vote is in. The rating is a Bradley-Terry Elo: strength converted with 400 × log₁₀(strength).
The chart in the thread shows the top 20. Krea 3 (Ultra Thinking) is the highlighted bar at 1298.
Design Arena's Image Editing leaderboard. Krea 3 (Ultra Thinking) is the highlighted bar at Elo 1298, between GPT Image 2 at 1347 and Reve 2.0 at 1295.
| Bar | Model, as labeled on the chart | Elo |
|---|---|---|
| 1 | GPT-Image-2.5 Sunburst | 1403 |
| 2 | GPT-Image-2.5 Flare | 1372 |
| 3 | GPT Image 2 | 1347 |
| 4 | Krea 3 (Ultra Thinking) | 1298 |
| 5 | Reve 2.0 | 1295 |
| 6 | Reve 2.1 | 1294 |
| 7 | ChatGPT Image | 1293 |
| 8 | Grok Imagine Image 2 A… (truncated on the chart) | 1276 |
| 9 | Muse Image | 1275 |
| 10 | Grok Imagine Image 2… (truncated on the chart) | 1274 |
| 11 | Seedream 5.0 Pro Edit | 1268 |
| 12 | MAI-Image-2.6 Flash | 1265 |
| 13 | GPT-Image-1.5 | 1253 |
| 14 | Gemini 3 Pro Image Preview 2K | 1252 |
| 15 | MAI-Image-2.6 | 1252 |
| 16 | Gemini 3 Pro Image Preview | 1245 |
| 17 | Ideogram 4.5 (High) | 1244 |
| 18 | Qwen Image 3 Pro | 1233 |
| 19 | MAI-Image-2.5 | 1231 |
| 20 | Grok Imagine | 1202 |
GPT Image 2 at 1347 sits directly above Krea 3. Reve 2.0 at 1295 sits directly below, 3 Elo back. Reve 2.1 is 1294 and ChatGPT Image is 1293. Two Grok labels are cut off on the chart, at 1276 and 1274.
What the thread shows it can do
The same thread walks through three edits.
1/ A reference, carried onto a model
Give Krea 3 a reference image and it carries that image onto a model with full fidelity, through dynamic poses, new camera angles, and shifting light. In the example, two sunglasses references and a portrait become the same person wearing those glasses under a new pose and a new light.
Reference to model. Two sunglasses references and a portrait become the same person wearing the glasses, through a new pose, a new camera angle, and shifting light.
2/ A fabric swatch, applied to any object
A fabric swatch can be applied to any object, and it follows how that object is built. Seams, cushions, and folds take the textile. That is what makes a product mockup look real. In the example, a rust corduroy and a blue-and-white stripe land on a beige sofa: rust on the cushions, stripe on the frame.
Fabric swatch on a sofa. Rust corduroy and a blue-and-white stripe follow the cushions, the frame, the seams, and the folds.
3/ A styled scene, built from a flat-lay
Krea 3 builds complete styled scenes with full outfits in moody, cinematic lighting. This one starts from a flat-lay of the clothes: a cream tee, dark jeans, and white sneakers. The result is three men on a rooftop at dusk, wearing that outfit.
A styled scene built from a flat-lay of a cream tee, dark jeans, and white sneakers. Three men on a rooftop at dusk, in that outfit, under moody cinematic light.
What that means for a creator
#4 of 51, 3 Elo above Reve 2.0 and 49 Elo behind GPT Image 2, with Krea as the #2 image-editing lab after OpenAI. The three examples are the practical version: a product reference that survives a new photo, a textile that follows the construction of a sofa, and a full outfit carried from a flat-lay into a lit scene. Architecture is #2. Landscape is #3.
How the three boards fit together
Someone looking for the best AI image editor will not get one rank out of this Krea 3 benchmark. The honest summary, as of October 19–20, 2026:
- On Arena’s multi-image board, Krea 3 Ultra is #4, tied on Arena’s rule with Nano Banana 2.1, and #2 in Commercial Design and 3D, on early vote counts.
- On Arena’s single-image board, Krea 3 Ultra is #8, with a rank spread of 5–10.
- On Artificial Analysis, Krea 3 Ultra is #5 and tied for #3 on overlapping intervals. Krea 3 Fast is #11. Ultra’s best categories are Animation & Gaming, Frontier, Composition & Framing, and Reasoning-Based Edit. Scene & Style Edit is not one of them.
- On Design Arena’s Image Editing leaderboard, Krea 3 (Ultra Thinking) is #4 of 51 at 1298 Elo, the #2 image-editing lab after OpenAI, ahead of Reve 2.0, #2 in Architecture and #3 in Landscape.
The pattern across Arena and Artificial Analysis is consistent even though the scales are not. Krea 3 Ultra is in the top group. GPT Image 2.5 leads the published top of both. On Arena’s single-image board, Grok Imagine Image 2.0, Nano Banana 2.1, and MAI-Image-2.6 also sit above Krea’s point rank. On Artificial Analysis, Nano Banana 2.1 and MAI-Image-2.6 share Ultra’s rank range. The intervals overlap the next cluster rather than clearing it. On Artificial Analysis, Ultra ranks higher where the edit has to keep a reference or reason about the scene than where it has to relight or restyle. Fast is a separate row there, not a footnote, and on two categories it ranks above Ultra.
If you want to try the model these pages measured, it is on krea.ai/krea-3. The product post, including what the model is for, is Introducing Krea 3.
Frequently asked questions
Ranks below are copied from the leaderboard pages and posts cited in this article. Each one is dated.
What is Krea 3's rank on the AI image editing leaderboards?
As of October 19–20, 2026, there are four answers. On Arena's Multi-Image Edit board, Krea 3 Ultra is #4 at 1435±11, preliminary, with 2,991 votes and a rank spread of 4–5 (page dated October 19). On Arena's single-image Image Edit board, Krea 3 Ultra is #8 at 1411±11, preliminary, with 2,619 votes and a spread of 5–10. On Artificial Analysis AA-Image-Editing v2.1, Krea 3 Ultra is #5 at 1151 Elo, rank range 3–6, and Krea 3 Fast is #11 at 1125 Elo (live page read October 20). On Design Arena's Image Editing leaderboard, Krea 3 (Ultra Thinking) is #4 of 51 at 1298 Elo, the #2 image-editing lab after OpenAI.
Is Krea 3 the best AI image editor?
No single board says that, and the boards do not share one rank. As of October 19–20, 2026, Krea 3 Ultra is #4 on Arena Multi-Image Edit, #8 on Arena single-image edit, #5 on Artificial Analysis (statistically tied for #3), and #4 of 51 on Design Arena. OpenAI's GPT Image 2.5 models lead the published tops of all three. Krea 3 Ultra's own best category ranks in this set are #2, on Arena's multi-image Commercial Design and 3D boards, on Design Arena's Architecture board, and on Artificial Analysis's Animation and Gaming and Frontier categories.
What does an AI image editing leaderboard measure?
All three sites rank models from blind pairwise human votes. A person sees two edits of the same image and the same instruction and picks one. Arena and Artificial Analysis turn those votes into Elo with a Bradley-Terry model and publish a confidence interval. Design Arena approximates Elo the same way. The scores are not on one scale: Artificial Analysis anchors FLUX.2 [dev] at 1000, and a 1435 on Arena is not comparable to a 1151 on Artificial Analysis or a 1298 on Design Arena.
How do Krea 3 Ultra and Krea 3 Fast differ on the Krea 3 benchmark?
Only Artificial Analysis lists both modes. As of the October 20, 2026 read, Ultra is #5 at 1150.58 Elo (displayed 1151) on 5,378 samples, and Fast is #11 at 1125.11 Elo (displayed 1125) on 5,087 samples. Fast's point Elo is above Nano Banana 2 (1123.64) and Nano Banana Pro (1106), with overlapping confidence intervals. Ultra is #4 on Reasoning-Based Edit and Fast is #20. Fast ranks above Ultra on Live-Action Film (#8 vs #10) and on Scene and Style Edit (#11 vs #18). Arena and Design Arena rank Krea 3 Ultra (Ultra Thinking), not Fast.
Where does Krea 3 rank on Design Arena?
Krea 3 (Ultra Thinking) ranks #4 of 51 on Design Arena's Image Editing leaderboard at 1298 Elo, which makes Krea the #2 image-editing lab after OpenAI. It edges out Reve 2.0, and ranks #2 in Architecture and #3 in Landscape. Design Arena posted the result on October 20, 2026.
Why is Krea 3 at 1435 on Arena and 1151 on Artificial Analysis?
They are different Elo scales, as of October 19–20, 2026. Arena's Multi-Image Edit score for Krea 3 Ultra is 1435±11. Artificial Analysis's AA-Image-Editing v2.1 score for Krea 3 Ultra is 1151 (1150.58 exact), on a scale anchored so FLUX.2 [dev] equals 1000. Design Arena's Image Editing chart prints 1298, on a scale whose docs describe a base of 1200. Compare ranks and intervals inside one board. Do not subtract the scores.
How many votes are behind these Krea 3 ranks?
On the Arena pages dated October 19, 2026, Krea 3 Ultra has 2,991 votes on Multi-Image Edit and 2,619 on Image Edit, and both rows are marked Preliminary. Category rows range from 229 to 774 votes. On Artificial Analysis, as of October 20, 2026, Ultra has 5,378 overall samples and Fast has 5,087. Category samples for Ultra run from 487 to 811. Design Arena's chart does not print a vote count for Krea 3.
Try Krea 3
The model measured on these leaderboards is live on Krea. Ultra and Fast are the modes in the Artificial Analysis results.
Try Krea 3
