1
GPT Image 2.5
OpenAI · Precise instruction-following and edits that leave the rest alone
$$$
GPT Image 2.5 Sunburst is #1 on five image arena boards: LMArena text-to-image, single-image edit and multi-image edit, and Artificial Analysis text-to-image and editing. Flare is the faster default and sits at #2 or #3 on the same boards at the same price.
- Cheapest way in
- Free tier in ChatGPT; paid plans raise limits; API pay as you go
- Cost per output
- API per 1024x1024 image: $0.006 low, $0.013 medium, $0.053 high, $0.094 xhigh, $0.211 max (plus input tokens)
- Reference images
- Multiple; count not published
- Character consistency
- Improved across successive edits (vendor claim)
- Editing
- Yes, including masked and multi-image
- Text rendering
- Strong; quote exact wording
- Max resolution
- Priced sizes up to 1536x1024
- Variants
- Flare (fast default), Sunburst (most capable)
Good at
- #1 across five arena boards
- Strong preservation of subjects across successive edits (OpenAI claim)
- Available in ChatGPT and via API
Watch out for
- Most expensive at max quality ($0.211 per 1024x1024)
- OpenAI has not published a reference-image limit for 2.5
- Arena scores still marked preliminary
Benchmarks LMArena (arena.ai) Image Edit Arena, single image: Sunburst #1, score 1517; Flare #2, 1482 · LMArena (arena.ai) Image Edit Arena, multi-image: Sunburst #1, score 1521; Flare #2, 1477 · Artificial Analysis Image Editing Arena (AA-Image-Editing v2.1): Sunburst #1, Elo 1199; Flare #3, 1182 · LMArena (arena.ai) Text-to-Image Arena: Sunburst #1, score 1427
Prices developers.openai.com, checked 1 November 2026
GPT Image 2.5 website Work made with it on AI LIVE
2
Nano Banana 2.1
Google · Character and product consistency across many references, at a low price
$
Nano Banana 2.1 (6 October 2026) is top five on LMArena multi-image edit and tied for 4th to 7th on Artificial Analysis editing, while costing about a sixth of GPT Image 2.5 at max quality. Google reports a big jump in multi-character consistency, and it accepts up to 14 references with 4K output.
- Cheapest way in
- Free image generation in the Gemini app; API is paid only
- Cost per output
- $0.0336 (1K), $0.0504 (2K), $0.113 (4K) per image; batch half price. Nano Banana Pro $0.134 (1K/2K), $0.24 (4K)
- Reference images
- Up to 14 (some sources say 10)
- Character consistency
- 1106 Elo vs 978 for Nano Banana 2 (Google's own test)
- Editing
- Conversational and mask-based
- Text rendering
- Improved, including infographics
- Max resolution
- 4K
Good at
- Up to 14 reference images reported (10 object, 4 character)
- 1K, 2K and 4K output
- $0.0336 per 1K image, half the price of Nano Banana 2
Watch out for
- Free Gemini app tier may still serve Nano Banana 2, not 2.1
- No free API tier for image output
- Some testers still prefer Nano Banana Pro's look
Benchmarks LMArena (arena.ai) Image Edit Arena, multi-image: #5, score 1432 (preliminary) · LMArena (arena.ai) Image Edit Arena, single image: #7, score 1427 (preliminary) · Artificial Analysis Image Editing Arena (AA-Image-Editing v2.1): #6 (tied with MAI-Image-2.6 and Krea 3 Ultra), Elo 1151 · Artificial Analysis Text to Image Arena (AA-Image-T2I v2.1): #4, Elo 1176
Prices ai.google.dev, checked 1 November 2026
Nano Banana 2.1 website Work made with it on AI LIVE
3
Seedream 5.0 Pro
ByteDance Seed · Multi-reference compositing of faces, products and objects
$$
Seedream 5.0 Pro is #6 on LMArena's multi-image edit board with over 100,000 votes, the most-tested model in the top ten, and #9 on Artificial Analysis editing. It accepts 10 to 14 references per request and is built into many creator platforms.
- Cheapest way in
- Pay per use via BytePlus ModelArk, resellers and platforms like Artlist
- Cost per output
- About $0.045 to $0.09 per image depending on provider
- Reference images
- Up to 14 (OpenRouter) or 10 (other sources)
- Character consistency
- Faces and products kept across references
- Editing
- Yes, image-to-image and multi-reference
- Text rendering
- Multilingual
- Max resolution
- 2K confirmed
Good at
- Strong multi-reference blending with a large, stable vote base
- Cheap extra references (about $0.003 each reported)
- Widely available via resellers and platforms
Watch out for
- 4K support disputed
- Reference limit differs by source
- Closed weights
Benchmarks LMArena (arena.ai) Image Edit Arena, multi-image: #6, score 1413 · LMArena (arena.ai) Image Edit Arena, single image: #13, score 1390 · Artificial Analysis Image Editing Arena (AA-Image-Editing v2.1): #9, Elo 1131
Prices atlascloud.ai, checked 1 November 2026
Seedream 5.0 Pro website Work made with it on AI LIVE
4
MAI-Image-2.6
Microsoft AI · Complex edits and campaign assets on a budget
$
MAI-Image-2.6 is #6 on LMArena single-image edit and tied #4 on Artificial Analysis editing, at about $0.04 per image. Access is mainly through Microsoft Foundry and API routers, which makes it more of a pipeline model than a creator app.
- Cheapest way in
- Pay-per-use via Microsoft Foundry or OpenRouter
- Cost per output
- About $0.039 per image (Artificial Analysis); tokens $5/M text in, $8/M image in, $38/M image out
- Reference images
- 5
- Editing
- Yes
- Text rendering
- Positioned as a strength (vendor)
- Max resolution
- Not stated
Good at
- Top-six on both editing arenas
- About $0.04 per image; Flash version roughly half that
- Up to 5 reference images; Bing web grounding
Watch out for
- Limited creator-facing apps (Foundry and API routers; Adobe Firefly announced as coming)
- Preview vs general availability status reported inconsistently
- Not on LMArena multi-image board top 30
Benchmarks LMArena (arena.ai) Image Edit Arena, single image: #6, score 1427 · Artificial Analysis Image Editing Arena (AA-Image-Editing v2.1): #4 (tied), Elo 1151 · LMArena (arena.ai) Text-to-Image Arena: #4, score 1329
Prices runtimewire.com, checked 1 November 2026
MAI-Image-2.6 website
5
Krea 3
Krea · Fast, targeted edits that keep products, logos and characters intact
$
Krea 3 (15 October 2026) debuted at #4 on LMArena multi-image edit and tied #5 on Artificial Analysis editing, which puts an independent lab level with Google and Microsoft. Most edits finish in under 8 seconds at native 2K with up to 6 references, though its arena vote counts are still small.
- Cheapest way in
- Krea free plan (100 compute units/day); Basic from $9/month ($5 billed yearly)
- Cost per output
- API $0.03 per 1K image, $0.06 per 2K image
- Reference images
- 6
- Character consistency
- Products and characters kept across images (vendor)
- Editing
- Yes, change-only-what-you-ask
- Text rendering
- Keeps existing text intact (#4 for text rendering, per Krea)
- Max resolution
- Native 2K
Good at
- Top-five editing on both arenas
- Up to 6 references per edit; keeps labels and logos
- $0.06 per 2K image, under 8 seconds
Watch out for
- Few arena votes so far (preliminary)
- Open weights not stated
- Weaker on scene and style edits (#18 in that Artificial Analysis category, per Krea)
Benchmarks LMArena (arena.ai) Image Edit Arena, multi-image: Ultra #4, score 1434 (preliminary) · LMArena (arena.ai) Image Edit Arena, single image: Ultra #9, score 1411 (preliminary) · Artificial Analysis Image Editing Arena (AA-Image-Editing v2.1): Ultra #5 (tied), Elo 1151; Fast #12, Elo 1126
Prices krea.ai, checked 1 November 2026
Krea 3 website Work made with it on AI LIVE
6
Grok Imagine Image 2.0
SpaceXAI · Cheap, capable single-image edits
$
Grok Imagine Image 2.0 is #5 on LMArena single-image edit, though only #17 on Artificial Analysis editing, so the two arenas disagree. At $0.04 per image it is good value for straightforward edits.
- Cheapest way in
- In the Grok app; $0.04 per image via the xAI API
- Cost per output
- $0.04 per image via xAI API
- Reference images
- Not documented
- Editing
- Yes
- Text rendering
- Not documented
- Max resolution
- Up to 2K via resellers; xAI does not state it
Good at
- #5 on LMArena single-image edit
- $0.04 per image official price
- Same model for generation and editing
Watch out for
- Much lower on Artificial Analysis editing (#17)
- Reference limit and max resolution not documented by xAI
- LMArena score preliminary
Benchmarks LMArena (arena.ai) Image Edit Arena, single image: #5, score 1439 (preliminary) · Artificial Analysis Image Editing Arena (AA-Image-Editing v2.1): #17, Elo 1117
Prices docs.x.ai, checked 1 November 2026
Grok Imagine Image 2.0 website
7
FLUX 3 Image
Black Forest Labs · Controllable editing you can self-host and fine-tune
$
FLUX 3 Image is #15 on Artificial Analysis editing, and the open-weights Klein model supports up to 10 references at native 4K for $0.009 per 1K image. For pipelines that need local control, LoRAs or ComfyUI, it is the strongest open option and replaces FLUX Kontext, which now ranks near the bottom of both edit arenas.
- Cheapest way in
- Pay as you go; Klein weights downloadable
- Cost per output
- $0.048 per image (FLUX 3 Image); from $0.009 per 1K image (Klein)
- Reference images
- Up to 10 (Klein)
- Character consistency
- Via multi-reference and LoRAs
- Editing
- Yes
- Text rendering
- Not documented
- Max resolution
- Native 4K (Klein)
- Open weights
- Klein only
Good at
- Open weights (Klein), commercial use free under $5M revenue
- Up to 10 references, native 4K (Klein)
- Very low API cost
Watch out for
- Behind closed leaders on editing arenas
- FLUX 3 not yet on LMArena edit boards (FLUX.2 [max] is #35)
- Flagship weights not public
Benchmarks Artificial Analysis Image Editing Arena (AA-Image-Editing v2.1): FLUX 3 Image #15, Elo 1122; Klein #27, Elo 1072 · LMArena (arena.ai) Image Edit Arena, single image: FLUX.2 [max] #35, score 1261; FLUX.1 Kontext [pro] #54
Prices bfl.ai, checked 1 November 2026
FLUX 3 Image website
8
Ideogram 4.5
Ideogram · Masked, repeatable design edits without colour or texture drift
$$
Ideogram 4.5 (30 September 2026) is built around Precise Edit, which changes only a masked area and is designed to avoid drift over many rounds. It is #14 on LMArena multi-image edit and #22 on single-image edit, and keeps Ideogram's typography strengths.
- Cheapest way in
- Free tier (30 points a week); Plus $20/month ($15 billed yearly)
- Cost per output
- API $0.008 to $0.22 per image depending on quality and mode
- Reference images
- Up to 4 (3 with a mask)
- Editing
- Precise Edit with mask or context window
- Text rendering
- Core strength
- Max resolution
- Native 2K (up to about 4.2MP)
Good at
- Mask-based Precise Edit with up to 4 references
- Good for swapping text on packaging and layouts
- Native 2K
Watch out for
- Edit jobs cost $0.22 at High
- Mid-table on Artificial Analysis editing (#28)
- Scores still preliminary on LMArena
Benchmarks LMArena (arena.ai) Image Edit Arena, multi-image: #14, score 1337 (preliminary) · LMArena (arena.ai) Image Edit Arena, single image: #22, score 1347 (preliminary) · Artificial Analysis Image Editing Arena (AA-Image-Editing v2.1): Ideogram 4.5 (High) #28, Elo 1070
Prices glbgpt.com, checked 1 November 2026
Ideogram 4.5 website
9
Qwen-Image-2.1 and Qwen-Image-Edit
Alibaba · Open-weights editing you can run yourself
$
Qwen-Image-2.1 is the top open-weights model on LMArena single-image edit (#20), and the older Qwen-Image-Edit-2511 is Apache-licensed and used by millions of LMArena voters (#41). Together they give a free, local route to instruction-based editing.
- Cheapest way in
- Free to download and run locally
- Cost per output
- Hosted Qwen Image Edit Plus 2511 about $0.06 per image (Artificial Analysis); closed Qwen-Image-3.0-Pro about $0.04
- Reference images
- Multi-image input (Edit 2511)
- Editing
- Instruction-based
- Max resolution
- 2048x2048 (2.1)
- Open weights
- Yes (2.1 research licence; Edit-2511 Apache 2.0)
Good at
- Open weights, ComfyUI support
- Qwen-Image-Edit-2511 permits commercial use (Apache 2.0)
- Native 2K with RGBA (2.1)
Watch out for
- Qwen-Image-2.1 is non-commercial (research licence)
- Well behind closed leaders on Artificial Analysis editing
- Newest Qwen-Image-3.0 is closed
Benchmarks LMArena (arena.ai) Image Edit Arena, single image: Qwen-Image-2.1 #20, score 1362 (preliminary); Qwen-Image-Edit (2511) #41, 1233 · Artificial Analysis Image Editing Arena (AA-Image-Editing v2.1): Qwen-Image-2.1 #24, Elo 1081; Qwen Image Edit Plus 2511 #52, Elo 991
Prices artificialanalysis.ai, checked 1 November 2026
Qwen-Image-2.1 and Qwen-Image-Edit website
10
Recraft V4.1
Recraft · Brand control: saved styles, colour palettes and editable vectors
$
Recraft's control is about brand rather than pixels: saved style IDs, up to 10 style references, explicit colour controls and SVG output that a designer can edit afterwards. It is not a leader on general edit arenas, so it ranks last here for the specific job of keeping a brand system consistent.
- Cheapest way in
- Free plan (no commercial rights); API from $0.035 per image
- Cost per output
- API $0.035 raster, $0.08 vector, $0.21 Pro raster, $0.30 Pro vector
- Reference images
- Up to 10 style references
- Character consistency
- Style consistency rather than character
- Editing
- Image-to-image
- Text rendering
- Not documented for V4.1
- Max resolution
- 2K (Pro)
- Vector
- Yes
Good at
- Editable SVG
- Saved styles and colour/background controls
- Low raster cost
Watch out for
- Editing limited to image-to-image in the V4.1 API
- Not on the top edit-arena boards
- Free plan has no commercial rights
Benchmarks LMArena (arena.ai) Text-to-Image Arena: V4.1 Utility Pro #26, score 1168 · Artificial Analysis Text to Image Arena (AA-Image-T2I v2.1): V4.1 Utility #29, Elo 1026
Prices recraft.ai, checked 1 November 2026
Recraft V4.1 website
