Best AI image generators (October 2026), according to LMArena and Artificial Analysis: OpenAI's GPT-Image-2.5 ranks first on both

The best AI for image generation, if you go by LMArena's and Artificial Analysis's public leaderboards, is OpenAI's GPT-Image-2.5 Sunburst as of October 23, 2026. It ranks first on LMArena's text-to-image leaderboard (snapshot dated October 7) and first on Artificial Analysis's text-to-image and image-editing leaderboards. "Best" here means "most preferred by those leaderboards' voters," not best at every job, and below OpenAI's models the two boards disagree. This report covers what each leaderboard shows, how it works, and what the leading image models cost and where you can use them.

This report is based on LMArena's and Artificial Analysis's published leaderboards and on each model maker's own pricing pages, model pages and announcements, checked October 23, 2026. Leaderboard positions are those evaluators' results, not 9to5AI's. For Black Forest Labs' open-weight option, see our FLUX 3 Image Klein story.

The short version

• Ranked first: GPT-Image-2.5 Sunburst, on LMArena's text-to-image board (1425, preliminary) and Artificial Analysis's text-to-image (Elo 1222) and editing (1199) boards.
• OpenAI holds the top three places by score on both text-to-image boards: Sunburst, Flare and GPT-Image-2.
• Next tier: Google's Nano Banana 2.1, Microsoft's MAI-Image-2.6 and xAI's Grok Imagine Image 2.0, in a different order on each board.
• Not on the boards: Midjourney appears on neither. Black Forest Labs' FLUX 3 Image is on Artificial Analysis (ranks 8–15) but not LMArena's text-to-image list.
• Price spread: from $0.01 an image (Meta's Muse Image API) to $210.7 per 1,000 images for GPT-Image-2.5 at its top setting, as Artificial Analysis lists it.

Which AI image generator ranks first on LMArena?

LMArena's text-to-image leaderboard, in its snapshot dated October 7, 2026, puts OpenAI's GPT-Image-2.5 Sunburst first at 1425 ±7, on 17,752 votes. It is marked "preliminary," as is the second-place GPT-Image-2.5 Flare (1398 ±7). The board covered 82 models and 6,604,125 votes.

LMArena text-to-image leaderboard, top 10

Arena score (higher means voters preferred it more often)

GPT-Image-2.5 Sunburst (OpenAI)1425

GPT-Image-2.5 Flare (OpenAI)1398

GPT-Image-2 Medium (OpenAI)1383

Grok Imagine Image 2.0 (xAI)1334

MAI-Image-2.6 (Microsoft AI)1332

Nano Banana 2.1 (Google)1328

Reve 2.1 (Reve)1302

Grok Imagine Image 2.0, Aug. 1 build (xAI)1294

Muse Image (Meta)1274

Reve 2.0 (Reve)1269

Show the data

Model	Score
GPT-Image-2.5 Sunburst (OpenAI)	1425
GPT-Image-2.5 Flare (OpenAI)	1398
GPT-Image-2 Medium (OpenAI)	1383
Grok Imagine Image 2.0 (xAI)	1334
MAI-Image-2.6 (Microsoft AI)	1332
Nano Banana 2.1 (Google)	1328
Reve 2.1 (Reve)	1302
Grok Imagine Image 2.0, Aug. 1 build (xAI)	1294
Muse Image (Meta)	1274
Reve 2.0 (Reve)	1269

Below OpenAI's three models (GPT-Image-2 Medium is third at 1383), the order is:

• 4. Grok Imagine Image 2.0 (xAI, listed by LMArena under "SpaceXAI"): 1334 ±12, preliminary, on only 3,113 votes.
• 5. MAI-Image-2.6 (Microsoft AI): 1332 ±5.
• 6. Nano Banana 2.1 (Google): 1328 ±9, preliminary. LMArena's changelog shows it was added on October 6.
• 7. Reve 2.1: 1302. 8. an August 1 build of Grok Imagine Image 2.0: 1294. 9. Muse Image (Meta): 1274. 10. Reve 2.0: 1269.

Ranks 4 to 6 are 6 points apart, and their error margins overlap, so their order isn't settled. Further down, Black Forest Labs' best entry is FLUX.2 [max] at 29th (1162); no FLUX 3 image model is listed. OpenAI's older GPT-Image-1.5 (High Fidelity) is 17th (1237). OpenAI's deprecations page schedules gpt-image-1.5 to leave the API on December 1, 2026, with the two GPT-Image-2.5 models as replacements, and lists gpt-image-1 for shutdown on October 23, 2026; see our November–December calendar.

How LMArena works: visitors type a prompt, see two anonymous images and pick one. Scores are a rating built from those votes. That captures broad visual preference, not text accuracy, edit precision or speed. "Preliminary" means a model doesn't yet have enough votes for a stable score. The page appears to apply style control by default but doesn't say so explicitly.

Which AI image generator ranks first on Artificial Analysis?

Artificial Analysis's text-to-image leaderboard (AA-Image-T2I v2.1), checked October 23, 2026, also puts GPT Image 2.5 Sunburst (max) first, at Elo 1222 (interval 1212–1232). The page lists each model's place alongside a rank range, which shows where models are statistically tied:

• 1: GPT Image 2.5 Sunburst (max), 1222
• 2: GPT Image 2.5 Flare (max), 1207
• 3 (range 3–4): GPT Image 2 (high), 1186
• 4 (range 3–6): Nano Banana 2.1, 1176
• 5 and 6 (range 4–6): MAI-Image-2.6, 1166; Grok Imagine Image 2.0, 1166
• 7 (range 7–8): Nano Banana 2 (Gemini 3.1 Flash Image), 1140
• 8 (range 7–12): Muse Image (Meta), 1127
• 13 (range 8–15): FLUX 3 Image (Black Forest Labs), 1117

How it works: Artificial Analysis says its Elo scores come from "blind preference votes by our validated private evaluators," a smaller, vetted pool rather than the open public. Its page shows release dates for each model but no date for the leaderboard itself. Its scale differs from LMArena's, so a 1222 here can't be compared with a 1425 there.

Which AI is best for editing images?

On Artificial Analysis's image-editing leaderboard (AA-Image-Editing v2.1), checked October 23, 2026, GPT Image 2.5 Sunburst (max) ranks first at Elo 1199, followed by Flare (max) at 1182. MAI-Image-2.6, Krea 3 (Ultra) and Nano Banana 2.1 follow at 1151 each, sharing ranks 3–6. Krea 3 launched October 15; see our Krea 3 story for its standard API prices.

LMArena's image-edit leaderboard, as published in LMArena's public leaderboard dataset on Hugging Face (snapshot dated October 20, 2026), also puts GPT-Image-2.5 Sunburst first, at 1514, ahead of Flare (1481) and GPT-Image-2 (Medium) (1461). MAI-Image-2.6 is sixth there (1425). Microsoft said on September 4 that MAI-Image-2.6 ranked second on LMArena's editing board at launch; that's a vendor statement about an older snapshot.

Why don't the two leaderboards agree?

They agree on the top: OpenAI's two GPT-Image-2.5 models, then GPT-Image-2. Below that, three differences matter:

• Who votes. LMArena uses open crowd votes; Artificial Analysis uses vetted evaluators.
• Sample size. Grok Imagine Image 2.0's fourth place on LMArena rests on 3,113 votes and a ±12 margin; on Artificial Analysis it shares ranks 4–6.
• Which models are entered. FLUX 3 Image is on Artificial Analysis but not LMArena. Midjourney is on neither, so neither board says anything about it.

Neither board tests a specific job such as product shots, posters with exact text or brand-consistent characters. Google's own model card for Nano Banana 2.1, for example, says small text and full-page text rendering are weak points.

What do the top AI image generators cost, and where can you use them?

Last verified: October 23, 2026

For a like-for-like view, Artificial Analysis lists what 1,000 images at 1024×1024 cost on each maker's API with default settings: $210.7 for GPT-Image-2.5 Sunburst or Flare at "max," $33.6 for Nano Banana 2.1, $38.9 for MAI-Image-2.6, $60.0 for Grok Imagine Image 2.0, $48.0 for FLUX 3 Image and $10.0 for Muse Image. Those are its calculations at its settings, and they match Google's, xAI's (medium quality), Meta's and BFL's own per-image prices. At those listed rates, GPT-Image-2.5 at its top setting costs about six times as much per image as Nano Banana 2.1 (a 9to5AI calculation).

OpenAI's launch post calls Images 2.5 "state-of-the-art" and says it cuts generation time by up to 50% compared with Images 2.0; those are vendor claims. Google released Nano Banana 2.1 on October 6; it was the only Gemini API launch between October 1 and 8, as noted in our Gemini 4 Argon story.

Which open-weight image model ranks highest?

Artificial Analysis's FAQ, checked October 23, 2026, names BFL's FLUX 3 Image Klein (Elo 1043), Qwen-Image-2.1 (1042) and Fibo 2.0 (1039) as its top open-weights text-to-image models, well below the proprietary leaders. Klein, released October 15, leads that group by one point, inside its ±12 interval. On LMArena's list, which labels licenses but doesn't say which models have downloadable weights, the highest entries not labeled proprietary are Alibaba's Qwen-Image-2.1, 19th at 1223 under a "qwen-research" license, and Ideogram 4.0 (Quality), 20th at 1205 under the "Ideogram Open Model" license; Nvidia's Cosmos3-Super-Text2Image (Agentic) is 27th at 1165 under OpenMDW-1.1. FLUX 3 Image Klein isn't on LMArena's text-to-image board.

Which should you pick?

What's confirmed, reported and uncertain

• Confirmed (official pages, October 23, 2026): API prices for GPT-Image-2.5, Nano Banana 2.1, Grok Imagine Image 2.0, Muse Image and FLUX 3 Image; Midjourney plan prices; where each model is available as stated by its maker.
• Third-party, with limits: both leaderboards' positions, each tied to its snapshot or check date and method. Four of LMArena's top six are preliminary.
• Vendor claims: OpenAI's "state-of-the-art" and speed claims; Microsoft's "best price-per-Elo"; Google's own evaluation scores.
• Uncertain: MAI-Image-2.6's API price; which Gemini app plans get Nano Banana 2.1; Grok Imagine Image 2.0's availability in the Grok app; how often either leaderboard will reshuffle as preliminary scores settle.