Hayz Labs · https://hayzlabs.com/reviews/top/image-models/
reviews · top 20 list · checked 18 Oct
Top 20 AI image models
Models that make pictures from a description, and the ones that edit them.
Best: GPT Image 2.5 Sunburst (100). Best free: Nano Banana 2.1 (73). 19 of 20 scored.
Ordered by our score, built from the Text to Image board in Arena's open leaderboard dataset. One row per model line, newest version only. Open the Table to see the number that placed each one and how it was scored.
- 1GPT Image 2.5 SunburstOpenAI
Top image model for quality and precise edits, about 21 cents a max-quality image. On a budget, use Nano Banana 2.1.
Best for Designers and developers who need the most accurate image editsYou give up About $0.21 per image at max quality; slower than Flare
10098–100Pay as you go - 2GPT Image 2.5 FlareOpenAI
Second on every image board, at Sunburst's price; OpenAI says it is faster. If waits are fine, Sunburst scores higher.
Best for Apps that need near-top images with shorter waitsYou give up Same token price as Sunburst, which scores higher on every board
9290–94Pay as you go - 3Grok Imagine Image 2.0xAI
Top five image quality from $0.04 an image
Best for Top-five quality through the API, also built into Grok and XYou give up Nano Banana 2.1 scores about the same (73) for less per image
7471–78Pay as you go - 4MAI-Image-2.6Microsoft
Microsoft's own image model, in preview on Microsoft Foundry
Best for Teams on Microsoft Foundry; the model behind Copilot and Bing imagesYou give up Still in preview; Muse Image costs about $0.01 an image but scores lower (59)
7473–75No public price - 5Nano Banana 2.1Google
Near-top image quality for about 3 cents an image, free to try in Gemini. Skip only if you need the very best edits.
Best for Strong images cheaply, or free to try in the Gemini appYou give up Behind both GPT Image 2.5 models on every board, most clearly on edits
7371–75Free - 6Reve 2.1Reve
Ranks just under the leaders in blind image votes
Best for Making and editing images, with a usable free daily tierYou give up Free tier trains on your data; API about $0.20 per image
6664–68No public price - 7Muse ImageMeta
Top 10 image quality for about a cent an image, free in Meta AI. Pick Nano Banana 2.1 if you want the higher rank.
Best for Lots of images on a small budget, free in Meta AIYou give up Ranks below Nano Banana 2.1 and both GPT Image 2.5 models; no data-use statement
5958–60Free - 8Seedream 5.0 ProByteDance
ByteDance's top image model, just outside the top ten
Best for 4K images and editing, also inside Dreamina and CapCutYou give up About $0.09 per image; Seedream 5.0 Lite is the better value pick
5554–56No public price - 9Qwen-Image-3.0-ProAlibaba
Low per-image price for a top 15 image model
Best for Dense text, infographics and Chinese characters for 3 to 4 centsYou give up Muse Image ranks higher on this list and costs less
5553–56Pay as you go - 10Nano Banana 2 LiteGoogle
Same per-image price as Nano Banana 2.1 but ranks lower; pick 2.1
Best for Bulk thumbnails and drafts from Google's APIYou give up Nano Banana 2.1 costs the same per image and ranks higher
5452–55Pay as you go - 11Nano Banana ProGoogle
Older Google model; Nano Banana 2.1 ranks higher for less
Best for Infographics and text-heavy layouts from GoogleYou give up About four times Nano Banana 2.1's price per image, and ranks lower
5352–54Pay as you go - 12Qwen-Image-2.1Alibaba
Best scoring image model you can download and run yourself, but its license bars commercial use. Fine for research.
Best for Researchers and hobbyists who want a strong image model offlineYou give up Licence bars commercial use; needs a capable GPU
4846–49Free - 13Ideogram 4.0Ideogram
Open weights under a commercial license from Ideogram
Best for Posters and logos with readable text, from weights you can downloadYou give up Ranks just below Qwen-Image 2.1, also free to run at home
4443–45No public price - 14Luma UNI-1.1 MaxLuma AI
Luma's image model; at the edge of the top 20 on blind votes
Best for Image making from Luma, at the edge of the top 20 in blind votesYou give up No price or review in our files yet
4140–42No public price - 15Recraft V4.1 Utility ProRecraft
Design-focused image model from Recraft; few votes so far
Best for Logos and icons as real editable vector files, not pixelsYou give up Weak on general photos; free plan images are public and owned by Recraft
3736–39No public price - 16Cosmos3-Super-Text2ImageNVIDIA
Open weights from Nvidia with commercial use allowed
Best for Businesses that want open image weights with commercial use allowedYou give up Ranks in the lower half of this list; no review yet
3736–38Price varies - 17HunyuanImage 3.0Tencent
Large open weights image model from Tencent; check the community license terms
Best for Teams that want a large open image model to run themselvesYou give up Community licence terms to check; ranks near the bottom of this list
3434–35Price varies - 18FLUX.2 [dev]Black Forest Labs
Local-only open model under a non-commercial license
Best for Local image workflows and fine-tuning on your own GPUYou give up Not free for business use; needs a GPU with lots of memory
3333–34Pay as you go - 19Seedream 5.0 LiteByteDance
Cheaper ByteDance model at the low end of this list
Best for Cheap batch and e-commerce imagesYou give up Muse Image ranks far higher and costs less per image
3232–33No public price - 20FLUX 3 ImageBlack Forest Labs
Top 10 image model that takes up to ten reference images, about 5 cents each. Plain prompts: Nano Banana 2.1 is cheaper.
Best for Designers who combine up to ten reference images or place items with boxesYou give up Trains on your prompts and outputs unless you email to opt out; no arena rank yet
Not scored yetPay as you go
| # | Name | Score | What placed it | Price | Best for | You give up |
|---|---|---|---|---|---|---|
| 1 | GPT Image 2.5 SunburstOpenAI · Closed · our review | 100likely 98–100 | Ranked #1 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1425. Our score: from blind votes (Arena). Report a mistake | API $5 text in, $8 image in, $30 image out per 1M tokens | Designers and developers who need the most accurate image edits | About $0.21 per image at max quality; slower than Flare |
| 2 | GPT Image 2.5 FlareOpenAI · Closed · our review | 92likely 90–94 | Ranked #2 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1398. Our score: from blind votes (Arena). Report a mistake | Not confirmed | Apps that need near-top images with shorter waits | Same token price as Sunburst, which scores higher on every board |
| 3 | Grok Imagine Image 2.0xAI · Closed | 74likely 71–78 | Ranked #4 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1334. Our score: from blind votes (Arena). Report a mistake | API $0.04 to $0.08 per image by size and quality | Top-five quality through the API, also built into Grok and X | Nano Banana 2.1 scores about the same (73) for less per image |
| 4 | MAI-Image-2.6Microsoft · Closed | 74likely 73–75 | Ranked #5 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1332. Our score: from blind votes (Arena). Report a mistake | Not confirmed | Teams on Microsoft Foundry; the model behind Copilot and Bing images | Still in preview; Muse Image costs about $0.01 an image but scores lower (59) |
| 5 | Nano Banana 2.1Google · Closed · our review | 73likely 71–75 | Ranked #6 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1328. Our score: from blind votes (Arena). Report a mistake | API $0.0336 per 1K image, $0.0504 2K, $0.113 4K | Strong images cheaply, or free to try in the Gemini app | Behind both GPT Image 2.5 models on every board, most clearly on edits |
| 6 | Reve 2.1Reve · Closed | 66likely 64–68 | Ranked #7 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1302. Our score: from blind votes (Arena). Report a mistake | Not confirmed | Making and editing images, with a usable free daily tier | Free tier trains on your data; API about $0.20 per image |
| 7 | Muse ImageMeta · Closed · our review | 59likely 58–60 | Ranked #8 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1274. Our score: from blind votes (Arena). Report a mistake | API $0.01 per image (served by Meta) | Lots of images on a small budget, free in Meta AI | Ranks below Nano Banana 2.1 and both GPT Image 2.5 models; no data-use statement |
| 8 | Seedream 5.0 ProByteDance · Closed | 55likely 54–56 | Ranked #11 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1255. Our score: from blind votes (Arena). Report a mistake | Not confirmed | 4K images and editing, also inside Dreamina and CapCut | About $0.09 per image; Seedream 5.0 Lite is the better value pick |
| 9 | Qwen-Image-3.0-ProAlibaba · Closed | 55likely 53–56 | Ranked #12 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1255. Our score: from blind votes (Arena). Report a mistake | API $0.034 per 1K image, $0.069 2K (US region) | Dense text, infographics and Chinese characters for 3 to 4 cents | Muse Image ranks higher on this list and costs less |
| 10 | Nano Banana 2 LiteGoogle · Closed | 54likely 52–55 | Ranked #14 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1250. Our score: from blind votes (Arena). Report a mistake | API about $0.0336 per 1K image | Bulk thumbnails and drafts from Google's API | Nano Banana 2.1 costs the same per image and ranks higher |
| 11 | Nano Banana ProGoogle · Closed | 53likely 52–54 | Ranked #15 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1248. Our score: from blind votes (Arena). Report a mistake | API $0.134 per 1K/2K image, $0.24 4K | Infographics and text-heavy layouts from Google | About four times Nano Banana 2.1's price per image, and ranks lower |
| 12 | Qwen-Image-2.1Alibaba · Open weights · our review | 48likely 46–49 | Ranked #17 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1223. Our score: from blind votes (Arena). Report a mistake | Free download (Qwen Research License) | Researchers and hobbyists who want a strong image model offline | Licence bars commercial use; needs a capable GPU |
| 13 | Ideogram 4.0Ideogram · Open weights | 44likely 43–45 | Ranked #18 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1205. Our score: from blind votes (Arena). Report a mistake | Not confirmed | Posters and logos with readable text, from weights you can download | Ranks just below Qwen-Image 2.1, also free to run at home |
| 14 | Luma UNI-1.1 MaxLuma AI · Closed | 41likely 40–42 | Ranked #20 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1189. Our score: from blind votes (Arena). Report a mistake | Not confirmed | Image making from Luma, at the edge of the top 20 in blind votes | No price or review in our files yet |
| 15 | Recraft V4.1 Utility ProRecraft · Closed | 37likely 36–39 | Ranked #24 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1169. Our score: from blind votes (Arena). Report a mistake | Not confirmed | Logos and icons as real editable vector files, not pixels | Weak on general photos; free plan images are public and owned by Recraft |
| 16 | Cosmos3-Super-Text2ImageNVIDIA · Open weights | 37likely 36–38 | Ranked #25 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1165. Our score: from blind votes (Arena). Report a mistake | Free download (OpenMDW-1.1, commercial use allowed) | Businesses that want open image weights with commercial use allowed | Ranks in the lower half of this list; no review yet |
| 17 | HunyuanImage 3.0Tencent · Open weights | 34likely 34–35 | Tied #32 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1150. Our score: from blind votes (Arena). Report a mistake | Free download (Tencent Hunyuan community license) | Teams that want a large open image model to run themselves | Community licence terms to check; ranks near the bottom of this list |
| 18 | FLUX.2 [dev]Black Forest Labs · Open weights | 33likely 33–34 | Ranked #35 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1146. Our score: from blind votes (Arena). Report a mistake | Free download, non-commercial license; no hosted API | Local image workflows and fine-tuning on your own GPU | Not free for business use; needs a GPU with lots of memory |
| 19 | Seedream 5.0 LiteByteDance · Closed | 32likely 32–33 | Ranked #37 in a public blind vote for images (arena.ai)What it rests onTheir figure: arena Text to Image 1138. Our score: from blind votes (Arena). Report a mistake | Not confirmed | Cheap batch and e-commerce images | Muse Image ranks far higher and costs less per image |
| 20 | FLUX 3 ImageBlack Forest Labs · Closed · our review | Not scored | Tied #12 in a public blind vote for images (Artificial Analysis)NotesTheir figure: AA Text to Image 1116. Not on arena T2I, so listed after board models. AA T2I #12 of distinct models (tied). Today's AA moves it down one step. Report a mistake | Not confirmed | Designers who combine up to ten reference images or place items with boxes | Trains on your prompts and outputs unless you email to opt out; no arena rank yet |
Nothing matches all of that. , or try everything.
How we ranked this
Order. Ordered by our score, built from the Text to Image board in Arena's open leaderboard dataset. One row per model line, newest version only.
How we score this list. Our score, built from Arena's blind votes on images made from text. The full rules are on How we score. Open "How it was scored" on any row to see the working.
What we did not do. We did not run any of these ourselves. Every number comes from the public source linked on its row, checked on 18 Oct 2026.
Spot a mistake? A wrong rank, price or link: tell us and we will check it against the source. Corrections are dated in News.
Reviews are drafted with help from AI models, including Anthropic's Claude. Model scores come only from outside leaderboards and independent evaluations, by a written rule.