whatmodel.sh / image / Nano Banana (Gemini 2.5 Flash Image)
Nano Banana (Gemini 2.5 Flash Image)
↗ share
+ contribute
Nano Banana (Gemini 2.5 Flash Image) is a closed-weights image model from Google, released August 2025. Licence: proprietary.
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of Nano Banana (Gemini 2.5 Flash Image), out of 197 with a score, then more from Google.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 37 |  FLUX.2 [flex] | 62±2 | api $0.050/image | ||
| 38 |  HunyuanImage 3.0-Instruct | 62±3 | api $0.090/image +local | ||
| 39 | Nano Banana (Gemini 2.5 Flash Image) | 62±1 | api $0.039/image | ||
| 40 |  Wan 2.7 Image | 62±2 | api $0.030/image | ||
| 41 | Imagen 4 Ultra | 61±2 | api $0.060/image | ||
| more from Google | |||||
| 5 | Nano Banana 2.1 | 81±2 | api $0.034/image | ||
| 9 | Nano Banana 2 (Gemini 3.1 Flash Image) | 76±1 | api $0.067/image | ||
| 10 | Nano Banana Pro (Gemini 3 Pro Image) | 76±1 | api $0.13/image | ||
$ what it's good at
text to image▂▅▅▃▄▄█▃▁▂▁▁ #30 of 126 · 57% of its score · 7 tests
Beats Wan 2.7 Image Pro, Cosmos3-Super-Text2Image here, though they rank higher overall. Trails ERNIE-Image here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LMArena text-to-image | 1150.3 ✎ | ▂▂▃▄▇█▄▂▄▂·▂ beats 45/75 | 22% |
| ServImage task acceptance (human payment decisions, commercial design tasks) | 29.9% ✎ | ▂▃··▂··▃█··▂ beats 6/11 | 10% |
| GenExam (relaxed overall, GPT-5 judge) | 55.9% ✎ | ▆▆·▂▃····▂▂█ beats 7/13 | 8.7% |
| T2I-ReasonBench (overall) | 89.5% ✎ | ▅▂█▂▂▂····▂▃ beats 13/13 | 7.0% |
| TIIF-Bench v3 (short prompts, TIIF-Evaluator) | 89.2% ✎ | ▂█····▄▄▄··▂ beats 10/11 | 5.4% |
| MMMG image generation (average over 15 tasks) | 65.4% ✎ | █··█▄▄·▄▄·▄▄ beats 8/9 | 3.4% |
| GenAI-Bench (VQAScore overall) | 79.0% ✎ | █···█·█····█ beats 1/3 | 0.7% |
image to image / editing▁▁·▃▂▃█▃▃▄▁▁ #29 of 62 · 37% of its score · 8 tests
Beats FLUX.2 [flex] here, though they rank higher overall. Trails Kling Image O1, Wan 2.7 Image here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| ServImage task acceptance (human payment decisions, commercial design tasks) | 29.9% ✎ | ▂▃··▂··▃█··▂ beats 6/11 | 10% |
| Arena image edit | 1293.2 ✎ | ▂▁▁▂▆▇█▄▆▅▁▂ beats 26/53 | 8.3% |
| CEdit-Bench-CN overall (G_O, GPT-4o judge) | 73.6% ✎ | ▃·········█▂ beats 2/7 | 5.3% |
| GEBench GUI state generation (English GE-Score) | 52.1% ✎ | ▄▄····▄▄█▄·█ beats 5/8 | 4.2% |
| CEdit-Bench-EN overall (G_O, GPT-4o judge) | 72.0% ✎ | ▄·▄···▄██··▄ beats 2/7 | 3.6% |
| UnicBench (EN, overall) | 78.8% ✎ | ▂····▂▄▂▄·█▂ beats 9/11 | 3.0% |
| CREval-Bench creative image editing (Overall, GPT-4o QA evaluator) | 81.3% ✎ | ▄█▄··█··▄·▄▄ beats 7/8 | 2.0% |
| OmniContext average (GPT-4.1 judge) | 78.4% ✎ | ▃▃▃··▃·····█ beats 5/6 | 1.0% |
other tests5 not in a task · 15% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| InfoEdit infographic editing (average success rate) | 3.5% ✎ | ▄··▄█····▄▄▄ beats 2/6 | 6.4% |
| EdiVal multi-turn editing overall (EdiVal-O after turn 3) | 56.2% ✎ | ▂·▂▃·▂·▃▂·▃█ beats 10/14 | 3.1% |
| PaintBench precise editing (mIoU, 20-task average) | 11.1% ✎ | ▃·····█▅▃·▃▅ beats 7/9 | 2.2% |
| VVRBench-Challenge accuracy (720 harder verifiable tasks) | 1.1% ✎ | ▄·▄···█·▄▄·▄ beats 1/6 | 2.0% |
| ViGoR-Bench zero-shot visual reasoning, image models (Result Avg) | 47.7% ✎ | ▂···▂·▂█▆▅·▂ beats 14/15 | 1.4% |
$ where to run it
Through an API
api · $0.039/image · ai.google.dev ↗ · price sourceGoogle Gemini Developer API Standard paid tier; up to 1024x1024 generated image output only. Input/text/thinking/tool charges additional. Scheduled shutdown 2026-10-02. Product: Gemini Developer API. API model: gemini-2.5-flash-image. Settings: tier=Standard paid, resolution=up to 1024x1024, batch=no, n=1. Price basis: provider-published approximate image OUTPUT cost. Additional/excluded charges: Input tokens, text/thinking output tokens and optional tools. Currency: USD.
$ how the score was worked out · LMArena leaderboard dataset 30%, ServImage: An Image Generatio… 10%, Meituan LongCat — LongCat-Ima… 9% and 13 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json