Image · Text to Image · July 26, 2026
Text to Image Leaderboard
Every major text-to-image model ranked by Elo from blind head-to-head human votes — no cherry-picked grids, no lab-supplied samples.
Models ranked
45
Open weights
10
Arena appearances
408,987
Prompt categories
5
Highlights
Higher is better
Arena Elo
Overall leaderboard · best model per lab
- 1GPT Image 2 (high)1338
- 2Reve 2.11301
- 3MAI-Image-2.51269
- 4Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image)1262
- 5Riverflow 2.01252
- 6HiDream-O1-Image-1.51246
- 7Seedream 5.0 Pro1239
- 8Cosmos3-Super-Text2Image (agentic)1218
- 9grok-imagine-image-quality1203
Higher is better
Open Weights
Downloadable checkpoints · best per lab
- 1Cosmos3-Super-Text2Image (agentic)1218
- 2HiDream-O1-Image-Dev-26041189
- 3Ideogram 4.0 (Quality)1172
- 4ERNIE Image1166
- 5Qwen Image Max 25121156
- 6Ideogram 4.0 Fast (Quality)1154
- 7FLUX.2 [dev]1152
Lower is better
Price per Image
USD per generation, default quality · cheapest per lab
- 1FLUX.2 [dev] Turbo$0.01
- 2ERNIE Image Turbo$0.01
- 3FLUX.2 [dev]$0.01
- 4Krea 2 Medium Turbo$0.01
- 5MAI-Image-2.5-Flash$0.02
- 6grok-imagine-image$0.02
- 7Qwen Image Max 2512$0.02
- 8Reve 2.1$0.02
- 9Seedream 4.0$0.03
Global leaderboard
Elo from blind head-to-head votes · All categories
| Rank | Model | Creator | Elo | 95% CI | Appearances | $/Image | Released |
|---|---|---|---|---|---|---|---|
| 1 | GPT Image 2 (high) | OpenAI | 1338 | ±7 | 21,060 | $0.21 | Apr 2026 |
| 2 | Reve 2.1 | Reve | 1301 | ±11 | 9,322 | $0.02 | Jul 2026 |
| 3 | MAI-Image-2.5 | Microsoft AI | 1269 | ±9 | 9,313 | $0.05 | Jun 2026 |
| 4 | Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) | 1262 | ±10 | 9,231 | — | Jun 2026 | |
| 5 | Nano Banana 2 (Gemini 3.1 Flash Image Preview) | 1261 | ±7 | 19,334 | $0.07 | Feb 2026 | |
| 6 | GPT Image 1.5 (high) | OpenAI | 1260 | ±8 | 12,172 | $0.13 | Dec 2025 |
| 7 | Riverflow 2.0 | Sourceful | 1252 | ±8 | 7,760 | $0.15 | Feb 2026 |
| 8 | HiDream-O1-Image-1.5 | HiDream | 1246 | ±8 | 12,374 | $0.08 | Jun 2026 |
| 9 | Reve 2.0 | Reve | 1242 | ±10 | 3,739 | $0.02 | Jun 2026 |
| 10 | Seedream 5.0 Pro | Bytedance | 1239 | ±10 | 8,118 | $0.09 | Jul 2026 |
| 11 | Nano Banana Pro (Gemini 3 Pro Image) | 1223 | ±7 | 11,470 | $0.13 | Nov 2025 | |
| 12 | Cosmos3-Super-Text2Image (agentic)Open | NVIDIA | 1218 | ±8 | 9,748 | — | May 2026 |
| 13 | MAI-Image-2.5-Flash | Microsoft AI | 1206 | ±9 | 4,817 | $0.02 | Jun 2026 |
| 14 | grok-imagine-image-quality | xAI | 1203 | ±7 | 15,425 | $0.05 | Apr 2026 |
| 15 | Recraft V4.1 Utility | Recraft | 1203 | ±8 | 5,388 | $0.04 | May 2026 |
| 16 | Recraft V4.1 Utility Pro | Recraft | 1199 | ±7 | 14,861 | $0.21 | May 2026 |
| 17 | FLUX.2 [max] | Black Forest Labs | 1194 | ±7 | 11,234 | $0.07 | Dec 2025 |
| 18 | Cosmos3-Super-Text2Image-4StepOpen | NVIDIA | 1192 | ±11 | 4,774 | — | Jul 2026 |
| 19 | Seedream 4.0 | ByteDance Seed | 1191 | ±7 | 16,772 | $0.03 | Sep 2025 |
| 20 | Krea 2 Medium | Krea | 1189 | ±7 | 14,687 | $0.03 | May 2026 |
| 21 | HiDream-O1-Image-Dev-2604Open | HiDream | 1189 | ±7 | 7,067 | — | May 2026 |
| 22 | Krea 2 Large | Krea | 1189 | ±7 | 15,062 | $0.06 | May 2026 |
| 23 | MAI-Image-2 | Microsoft AI | 1187 | ±7 | 7,661 | $0.04 | Mar 2026 |
| 24 | FLUX.2 [pro] | Black Forest Labs | 1185 | ±7 | 10,525 | $0.03 | Nov 2025 |
| 25 | Krea 2 Medium Turbo | Krea | 1185 | ±11 | 2,941 | $0.01 | Jun 2026 |
| 26 | Imagen 4 Ultra Preview 0606 | 1180 | ±8 | 5,871 | $0.06 | Jun 2025 | |
| 27 | grok-imagine-image | xAI | 1179 | ±8 | 9,669 | $0.02 | Jan 2026 |
| 28 | FLUX.2 [flex] | Black Forest Labs | 1178 | ±8 | 6,350 | $0.06 | Nov 2025 |
| 29 | Luma UNI 1 Max | Luma Labs | 1173 | ±8 | 5,665 | $0.10 | May 2026 |
| 30 | Qwen Image 2.0 Pro (2026-04-22) | Alibaba | 1173 | ±8 | 5,698 | $0.07 | Apr 2026 |
| 31 | Ideogram 4.0 (Quality)Open | Ideogram | 1172 | ±8 | 9,937 | $0.10 | Jun 2026 |
| 32 | Imagen 4 Ultra | 1171 | ±8 | 6,705 | $0.06 | Jun 2025 | |
| 33 | ImagineArt 2.0 | ImagineArt | 1168 | ±7 | 10,839 | $0.03 | Apr 2026 |
| 34 | Seedream 4.5 | ByteDance Seed | 1168 | ±7 | 12,578 | $0.04 | Dec 2025 |
| 35 | ERNIE ImageOpen | Baidu | 1166 | ±9 | 4,727 | $0.03 | Apr 2026 |
| 36 | ERNIE Image TurboOpen | Baidu | 1162 | ±8 | 5,626 | $0.01 | Apr 2026 |
| 37 | Imagen 4 Preview 0606 | 1161 | ±8 | 5,766 | $0.04 | Jun 2025 | |
| 38 | MAI-Image-2-Efficient | Microsoft AI | 1156 | ±7 | 8,284 | $0.02 | Apr 2026 |
| 39 | Qwen Image Max 2512Open | Alibaba | 1156 | ±8 | 5,754 | $0.02 | Dec 2025 |
| 40 | Recraft V4.1 Pro | Recraft | 1155 | ±8 | 4,983 | $0.21 | May 2026 |
| 41 | Ideogram 4.0 Fast (Quality)Open | Fal | 1154 | ±10 | 3,518 | $0.02 | — |
| 42 | FLUX.2 [dev] TurboOpen | Fal | 1153 | ±8 | 6,315 | $0.01 | Dec 2025 |
| 43 | FLUX.2 [dev]Open | Black Forest Labs | 1152 | ±7 | 9,442 | $0.01 | Nov 2025 |
| 44 | Nano Banana (Gemini 2.5 Flash Image) | 1152 | ±7 | 11,088 | $0.04 | Aug 2025 | |
| 45 | Recraft V4.1 | Recraft | 1152 | ±8 | 5,317 | $0.04 | May 2026 |
Showing 45 of 45 models. Category Elo is computed from votes on that prompt type alone.
Arena Elo — Open Weights
Every downloadable checkpoint on the board, ranked by overall Elo. Higher is better.
Cosmos3 sits fifth overall — ahead of most closed APIs — and the FLUX.2 [dev] line keeps Black Forest Labs in four of the fourteen open slots.
Arena Elo vs. Price per Image
Crowd-judged quality against USD per generation (log scale).
Up and to the left wins. The open-weights mid-table — FLUX.2 [dev], Qwen-Image-2512, Hunyuan — is where the value lives; models without public per-image pricing are excluded.
Frontier Elo Over Time
Arena Elo by release date. The dashed line tracks the running frontier.
The frontier has changed hands five times since Imagen 4 Ultra Preview 0606 — Google, ByteDance Seed, and OpenAI have all held it. GPT Image 2 set the current mark in April 2026.
How the arena works
1 · Prompt
Prompts are sampled across five categories — people, photorealism, art, product, and typography — so no model can farm a single style.
2 · Blind matchup
Two models generate from the same prompt. Voters see the images side by side with no names, no logos, and no way to tell who made what.
3 · Elo
Each vote updates both models' Elo ratings. Confidence intervals come from resampling the vote history; new models stay provisional until they clear roughly 3,000 appearances.
FAQ
Which model leads the arena right now?
GPT Image 2 (high) holds the top spot with an Elo of 1338 across 21,060 appearances — a 37-point gap over the chasing pack.
And among open weights?
NVIDIA's Cosmos3-Super-Text2Image (agentic) leads the open-weights cut at 1218 — twelfth overall, ahead of plenty of closed APIs — followed by Cosmos3-Super-Text2Image-4Step (1192) and HiDream's HiDream-O1-Image-Dev-2604 (1189).
How are models ranked?
Pure crowd judgment. Voters get a prompt and two unlabeled images from different models and pick the better one; thousands of those blind matchups feed an Elo rating, with a 95% confidence interval from resampling the vote history.
What counts as open weights?
The checkpoint has to be downloadable and runnable on your own hardware under a published license. API-only siblings of open families — FLUX.2 [pro], Qwen Image 2.0 Pro — stay in the proprietary column.
Latest Research
Reporting from the paper mines
Essex County helps build New Jersey's innovation economy
By NJBIZ · 2 hrs ago
The new premium product: books written by people
By Financial Times News · 2 hrs ago
The People Most Likely To Resist AI, And Why They Resist
By Forbes · 6 hrs ago
Symposium "Building AI Readiness: Safety, Talent, and International Cooperation" | The University of Tokyo
By 東京大学 · 1 hr ago
Methodology: rankings come from blind pairwise human votes; Elo, confidence intervals, and appearance counts reflect the full vote history at snapshot time. Arena data: public image-arena snapshot, July 26, 2026. Prices are list API prices per image at default quality and change frequently.