Skip to main content

142 image models, 108 with our own test

Compare AI image models

Muse Image, Grok Imagine Image 2.0 and FLUX 3 Image are practically tied at the top of GraduallyBench Image v2, our own benchmark. In user votes, GPT Image 2.5 Sunburst leads. Cheap and good: Muse Image at $0.01 per image.

GraduallyBench Image v2Our own benchmark

90out of 100

Muse Image

practically tied: Grok Imagine Image 2.0 (90) and FLUX 3 Image (88)

Arena rating

1,429from user votes

GPT Image 2.5 Sunburst

then GPT Image 2.5 Flare (1,398) and GPT Image 2 (1,382)

Cheap and good

$0.01per image via API

Muse Image

rank 1 in GraduallyBench Image v2, cheapest of the top ten

What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.

The best AI image models in GraduallyBench Image v2

Our own benchmarkGraduallyBench Image v2 is our own benchmark: every model gets the same four image tasks, and two AI judges from different vendors compare the images in pairs without knowing the models. That gives models without an arena rating a comparable score too.

GraduallyBench Image v2

Mean of four fixed image tasks, judged pairwise and blind by two AI judges, 0 to 100 · higher is better

20 of 108 models

As of November 7, 2026 · Source: own test by gradually.ai

The best AI image models on LMArena

Text to image

Images from a text prompt, rated in anonymous LMArena blind comparisons · higher is better

20 of 21 models

As of November 13, 2026 · Source: LMArena

Image to image

Editing an existing image with a prompt, rated in anonymous LMArena blind comparisons · higher is better

20 of 21 models

As of November 13, 2026 · Source: LMArena

All images from GraduallyBench Image v2 at a glance

142 of 142 models

Number under the image: points in the task (0 to 100)Trophy and violet frame: best image of the taskColumns from left to right: Typography and layout, Product photography, Character and detail, Infographic

  1. 1Muse Image90LMArena: Rank 7 of 21

    $0.01

    per image
    100 images: $1.00

  2. 2Grok Imagine Image 2.090LMArena: Rank 5 of 21

    $0.06

    per image
    100 images: $6.00

  3. 3FLUX 3 Image88LMArena: not listed

    n/a

    no price known

  4. 4Nano Banana 2.186LMArena: not listed

    n/a

    no price known

  5. 5GPT Image 2.5 Sunburst86LMArena: Rank 1 of 21

    $0.21

    per image
    100 images: $21.07

  6. 6FLUX 3 Image Klein85LMArena: not listed

    n/a

    no price known

  7. 7GPT Image 2.5 Flare85LMArena: Rank 2 of 21

    $0.21

    per image
    100 images: $21.07

  8. 8GPT Image 284LMArena: Rank 3 of 21

    $0.21

    per image
    100 images: $21.10

  9. 9Seedream 5.0 Pro82LMArena: Rank 11 of 21

    $0.09

    per image
    100 images: $9.00

  10. 10Wan 2.7 Image Pro81LMArena: not listed

    n/a

    no price known

What does your image volume cost?

Enter how many images you create per month and narrow the list with the filters. Based on the API price per image, without subscription discounts.
28 of 142 models
28 of 142 models have a representative price per image. The switch “Show models without a price” adds the other 114, without a monthly cost.
Per month
Muse ImageJul 2026$0.011,26690$10.00
FLUX.2 Klein 9BOpen weightsconfiguration 9BJan 2026$0.0151,07052$15.00
Krea 2 Medium Turboconfiguration Medium TurboJun 2026$0.015n/a72$15.00
MAI-Image-2.6 FlashPreviewJul 2026$0.019n/a72$19.46
MAI-Image-2.5 FlashPreviewconfiguration FlashJun 2026$0.02n/a74$19.97
FLUX.2 Proconfiguration ProNov 2025$0.031,15362$30.00
Wan 2.6 Text to Imageconfiguration Text to ImageJan 2026$0.03n/a55$30.00
Nano Banana 2 Liteconfiguration Gemini 3.1 Flash Lite ImageJun 2026$0.0341,24956$33.60
Recraft V4.1 Utilityconfiguration UtilityMay 2026$0.035n/a68$35.00
MAI-Image-2.6PreviewJul 2026$0.0391,33066$38.91
HiDream-O1-ImageOpen weightsMay 2026$0.041,11440$40.00
MAI-Image-2.5PreviewJun 2026$0.0481,25372$48.13
Grok Imagine Image QualityOutdatedApr 2026$0.05n/a67$50.00
FLUX.2 Flexconfiguration FlexNov 2025$0.061,15666$60.00
Grok Imagine Image 2.0Aug 2026$0.061,31890$60.00
Ideogram 4.0Open weightsconfiguration StandardJun 2026$0.06n/a76$60.00
Krea 2 Largeconfiguration LargeMay 2026$0.061,11474$60.00
Nano Banana 2configuration Gemini 3.1 Flash Image PreviewFeb 2026$0.0671,25661$67.00
FLUX.2 Maxconfiguration MaxDec 2025$0.071,16167$70.00
Qwen Image 2.0 Proconfiguration 2026-04-22Apr 2026$0.075n/a76$75.00
HiDream-O1-Image-1.5Jun 2026$0.08n/a72$80.00
Seedream 5.0 ProJul 2026$0.091,24882$90.00
Ideogram 4.5Weights unknownn/a$0.10n/a58$100
GPT Image 1.5Outdatedconfiguration HighDec 2025$0.131,23671$133
Nano Banana Proconfiguration Gemini 3 Pro ImageNov 2025$0.131,24669$134
GPT Image 2.5 FlareSep 2026$0.211,39885$211
GPT Image 2.5 Sunburstconfiguration SunburstSep 2026$0.211,42986$211
GPT Image 2configuration HighApr 2026$0.211,38284$211

28 of 142 models have a representative price per image. The switch “Show models without a price” adds the other 114, without a monthly cost.

Three common questions, three lists

Cheap and good

60+ GraduallyBench Image v2 points, sorted by price per image

  1. 1.Muse Image$0.01 · 90
  2. 2.Krea 2 Medium Turbo$0.015 · 72
  3. 3.MAI-Image-2.6 Flash$0.019 · 72
  4. 4.MAI-Image-2.5 Flash$0.02 · 74
  5. 5.FLUX.2 Pro$0.03 · 62
  6. 6.Recraft V4.1 Utility$0.035 · 68
  7. 7.MAI-Image-2.6$0.039 · 66
  8. 8.MAI-Image-2.5$0.048 · 72
  9. 9.Grok Imagine Image Quality$0.05 · 67
  10. 10.FLUX.2 Flex$0.06 · 66

Best arena quality under $0.07

Sorted by LMArena rating

  1. 1.MAI-Image-2.61,330 · $0.039
  2. 2.Grok Imagine Image 2.01,318 · $0.06
  3. 3.Muse Image1,266 · $0.01
  4. 4.Nano Banana 21,256 · $0.067
  5. 5.MAI-Image-2.51,253 · $0.048
  6. 6.Nano Banana 2 Lite1,249 · $0.034
  7. 7.FLUX.2 Flex1,156 · $0.06
  8. 8.FLUX.2 Pro1,153 · $0.03
  9. 9.HiDream-O1-Image1,114 · $0.04
  10. 10.Krea 2 Large1,114 · $0.06

Open weights

Self-hostable, sorted by API price per image

  1. 1.FLUX.2 Klein 9B$0.015
  2. 2.HiDream-O1-Image$0.04
  3. 3.Ideogram 4.0$0.06
  4. 4.Anima Base v1.0n/a
  5. 5.AuraFlow v0.3n/a
  6. 6.BAGEL-7B-MoTn/a
  7. 7.Boogu-Image-0.1-Basen/a
  8. 8.Boogu-Image-0.1-Turbon/a
  9. 9.Chroma1-Flashn/a
  10. 10.Chroma1-HDn/a

142 important models and configurations

Quality tiers such as Max, Pro or High count as separate models when they have their own results and prices.

For the full technical profile of every model, browse the AI model directory.

Image model or AI image generator?

An image model is the underlying generation system, such as GPT Image 2.5 Sunburst, Midjourney V8.2, or Nano Banana 2. A generator is the app around one or more models, including its interface, subscription, editing workflow, storage, and usage rights.

Use this page to compare model performance. Use the generator guide to choose a complete product.

Frequently asked questions about the image model comparison

How GraduallyBench Image v2, arena coverage, prices, sample images, and data freshness work.

It compares separate arena results, task-specific prices, technical specifications, availability, and documented sample images. GraduallyBench Image v2 averages only the 4 fixed tasks. Research benchmarks and preference arenas remain separate.

GraduallyBench Image v2 uses 4 fixed tasks: Typography and layout, Product photography, Character and detail, and Infographic. Every model receives the same English prompt in exactly one request, and the first image counts, without selection or retouching. That is n = 1 per model and task. Two judge models from different vendors, Gemini 4 Argon (High) and GPT-6 Astra, compare the images in pairs without knowing the models, each pair in both screen orders. From all pair results, a Bradley-Terry model estimates each model's chance to beat an average model of the test cohort; that chance is the 0 to 100 points, with a 95 % range from a bootstrap. The overall score is the equal-weight mean of the 4 tasks. Judge models can have visual preferences of their own.

LMArena ranks models by human preference from anonymous image matchups. GraduallyBench Image v2 uses two automated judges on four fixed tasks, with one archived first image per model and task. They measure different things, so the rankings stay separate. The rankings on the image wall and in the cost calculator keep both sources apart.

The price per image is the representative value from the Artificial Analysis text-to-image ranking for the configuration tested there, converted from 1,000 images to one image. It is the same value the model profiles show. Resolution, quality tier, and subscription discounts can change the actual price. 114 of 142 models have no such price and are left out of the cost calculator.

No. All 10011 matchup pages follow the same evidence rules, but a page shows only measurements shared by the exact two checkpoints. GraduallyBench Image v2 currently covers 108 of 142 models with four archived first outputs. New checkpoints remain unranked until they complete the same protocol.

477 of 10,011 matchups have at least one shared measurement series in the stored snapshot. 9,515 of the 9,534 gaps involve Midjourney V8.2, Seedream 4.0, Luma Uni-1.1 Max, Firefly Image Model 5, Midjourney V8.1, FLUX 3 Image, FLUX 3 Image Klein, Nano Banana 2.1, Hy Image 3.5 Preview, Ming-Image-0.1-Design, Recraft V4.1, Recraft V4.1 Flash, Bria FIBO 2, Krea 3, Qwen Image 3.0 Pro, Qwen Image 3.0, Qwen Image 2.1 Pro, Qwen Image 2.1 Turbo, Wan 2.7 Image Pro, Wan 2.7 Image, Seedream 5.0 Flash, SenseNova U1.5 Lite, Kling Image 3.0, Kling Image 3.0 Omni, Vidu Q2 Image, WAND-Vega-Image 1.0 Pro, WAND-Vega-Image 1.0 Flash, WAND-Vega-Image 1.0 Lite, GLM-Image, MiniMax Image-01, Z-Image Turbo, Luma Uni-1.1, Runway Gen-4 Image, Runway Gen-4 Image Turbo, Leonardo Phoenix 1.0, Leonardo Lucid Origin, Leonardo Lucid Realism, Adobe Firefly Image 3, Adobe Firefly Image 4, Adobe Firefly Image 4 Ultra, Stable Diffusion 3.5 Medium, Stable Diffusion 3.5 Large Turbo, Stable Diffusion 3.5 Flash, Xingliu Star-3, MiracleVision V6, Hy Image 3.0, Bria FIBO Lite, Mystic 2.5, Mystic 2.5 Flexible, Mystic 2.5 Fluid, Recraft V4.1 Utility Pro, Seedream 4.5, Seedream 5.0 Lite, Kling Image O1, Stable Image Core, Stable Image Ultra, FLUX.2 Klein 9B Preview, GPT Image 1 Mini, ChatGPT Image Latest (API-Alias), Nano Banana (Gemini 2.5 Flash Image), FLUX.2 Klein 4B, FLUX.2 Dev, FLUX1.1 [pro], FLUX1.1 [pro] Ultra, FLUX.1 [dev], FLUX.1 Kontext [pro], FLUX.1 Kontext [max], Stable Diffusion XL 1.0, Stable Diffusion 3 Medium, SDXL Turbo, Stable Diffusion Turbo, Midjourney V7, Niji 7, Grok Imagine Image 1.0, HUMAIN Image 1, MOSS-Image-Base, MOSS-Image-SFT, LLaDA-Image, LLaDA-Image-Turbo, Nucleus-Image, LongCat-Image, BAGEL-7B-MoT, OmniGen2, Emu3.5-Image, Kandinsky-5.0-T2I-Lite-SFT, Kandinsky-5.0-T2I-Lite-Pretrain, NextStep-1.1, Boogu-Image-0.1-Base, Boogu-Image-0.1-Turbo, SANA-1.5 1.6B 1024px, SANA-1.5 4.8B 1024px, Sana-Sprint 0.6B 1024px, Sana-Sprint 1.6B 1024px, Chroma1-HD, Chroma1-Flash, AuraFlow v0.3, PixArt-Sigma-XL-2-1024-MS, Lumina-Image 2.0, HART-0.7B-1024px, Janus-Pro-1B, Janus-Pro-7B, Show-o2-1.5B-HQ, Kandinsky 3.1, Anima Base v1.0, Krea 2 Raw, HunyuanImage 3.0, HunyuanImage 3.0 Instruct Distil, Z-Image Base, GPT-Image-2.6 Sunburst, or GPT-Image-2.6 Flare. These models have no stored arena result. The remaining 19 pairs have results, but not from the same arena and category.

No. Prompts and selection rules are standardized, but provider resolution, quality tiers, prompt processing, and randomness can still differ. Some older samples lack original files and complete provider responses, so their generation history cannot be independently reconstructed in full.

Max, Pro, High, and other tiers can have different scores, prices, and runtime settings. Combining them would hide meaningful differences.

The stored LMArena snapshot was retrieved on November 13, 2026. GraduallyBench Image v2 was evaluated on November 7, 2026. Most priced Artificial Analysis text-to-image entries were retrieved on November 12, 2026. The Grok Imagine Image 2.0 price entry was retrieved on November 12, 2026. Model facts and sample images carry separate source dates.

Changelog

The latest updates and improvements to our image model comparison
November 7, 2026v1.9

GraduallyBench Image v2 with pairwise judging

  • GraduallyBench Image v2 now compares the same images in pairs, in both screen orders, with two judge models from different vendors; scores show the chance to beat an average model, with a 95 % range
November 2, 2026v1.8

Popular comparisons by demand

  • Popular comparisons and the further duels on every pair page now follow the computed comparison popularity, with each model's predecessor and successor
October 27, 2026v1.7

Image wall instead of separate leaderboards

  • Replaced the leaderboards, gallery matchup, price-quality map, and featured matchups with one image wall: every model, all four tasks, points and rank under each image, price per image
  • Added three short answer cards for GraduallyBench, the arena rating, and the cheapest good model with its subscription options
  • Opened every image in a large view with its original prompt
Show older versions (7)
October 17, 2026v1.6

Gallery matchup, price-quality map, and GraduallyBench

  • Placed the LMArena and GraduallyBench leaderboards side by side
  • Added a gallery matchup and an image wall with the archived first outputs of all four tasks
  • Added a price-quality map, a cost calculator, and three budget answers
  • Renamed the own image test to GraduallyBench and turned popular matchups into cards with scores and prices
September 22, 2026v1.5

35-model roster and re-verified pricing

  • Expanded the catalog to 35 models and 595 head-to-head matchups
  • Re-verified provider prices and lifecycle notices against current provider pages
September 19, 2026v1.4

GPT Image 2.5 Sunburst and Midjourney V8.2

  • Expanded the catalog to 30 models and 435 head-to-head matchups
  • Added exact LMArena results for GPT Image 2.5 Sunburst where available
  • Kept GPT Image 2 and Midjourney V8.1 samples under their original checkpoints
August 28, 2026v1.3

Complete image model leaderboard

  • Added an equal-weight leaderboard for all 28 comparison models
  • Combined only the four shared GraduallyBench tasks
  • Kept the n = 1 GraduallyBench ranking separate from research benchmarks and preference arenas
August 18, 2026v1.2

Complete matchup coverage and research benchmarks

  • Added four shared Gradually rankings from three model-blind orderings of all 28 first outputs
  • Integrated exact model results from GenExam, Qwen Image Bench, and WISE Verified
  • Kept n = 1 GraduallyBench rankings, academic benchmarks, and preference arenas visibly separate
August 18, 2026v1.1

FAQ, changelog, and internal links

  • Added an FAQ that quantifies shared arena coverage and explains data gaps
  • Added this changelog so material updates remain visible
  • Added the hub and all matchup pages to navigation, the sitemap, and related comparisons
August 17, 2026v1.0

Initial release

  • Published 28 current image models and 378 head-to-head matchups
  • Kept text-to-image generation and image editing in separate benchmark series
  • Combined pricing, specifications, sources, and documented sample images