On October 27 Microsoft shipped MAI-Image-2.7-Preview, and AI CEO Mustafa Suleyman's claim was blunt: MAI-Image is now the world's second-ranked image editing model. On Artificial Analysis's AA-Image-Editing v2.1 board (snapshot 2026-10-22) it sits at 1193 Elo, six points behind the top, GPT Image 2.5 Sunburst (max) at 1199, and ahead of Google, Meta, xAI, and OpenAI's other tier, GPT Image 2.5 Flare (1182, third). The catch: no API, no pricing, closed weights -- only three doors in, the MAI Playground, Arena, and Artificial Analysis. A price you can buy against a score you cannot: the two columns of this review.
A scope note: this piece covers AI image editing -- an existing picture plus an instruction, changing style, text, regions, or identity -- not OCR text recognition, which extracts characters already in the picture; the two are constantly confused and select tools differently.
Three leaderboards, three number ones
The roster, seven entries: Microsoft's MAI-Image-2.7-Preview and the on-sale MAI-Image-2.6, Google's Nano Banana 2.1 Edit, OpenAI's GPT Image 2, Meta's Muse Image Edit, and three Chinese contenders -- ByteDance's Jimeng (Seedream 5.0, internationally Dreamina), Alibaba's Tongyi Wanxiang, and Kuaishou's Kling.
Ground rules first: the three leaderboards cited here do not interoperate. ModelGrep's Image Editing Arena (Updated October 2026, 77 ranked models) tops with Nano Banana 2.1 Edit at 1151. AA-Image-Editing v2.1 (Artificial Analysis, snapshot 2026-10-22, figures relayed through Microsoft's launch charts) tops with GPT Image 2.5 Sunburst (max) at 1199. Arena Image Edit (LMArena caliber, October 20 snapshot) tops with GPT Image 2.5 Sunburst at 1514 (Q2 score). Three tops spread more than four hundred points apart -- three calibers, not three opinions; no two numbers compare across boards, and converting 2.7's 1193 into another ranking is an error.
The ModelGrep top five, October 2026 caliber: Nano Banana 2.1 Edit 1151, GPT Image 2 API 1136, Meta Muse Image Edit 1129, MAI Image 2.5 1127, MAI Image 2.5 Pro (Edit) 1123. Microsoft's highest entry here is the 2.5 (fourth); 2.7 is nowhere on this board. The Arena board carries two warnings: it is a preference vote that never verifies instruction-following, and three of the six models released in the past 60 days (FLUX 3 Image, Hy Image 3.5 Preview, MAI-Image-2.6 Flash) still have no score; Seedream 4.0 has dropped off, last recorded 1270 on the October 6 caliber.
One table, all seven
| Model | Vendor | ModelGrep, Oct 2026 | AA v2.1, snapshot 2026-10-22 | API | Price (standard tier) | One-liner |
|---|---|---|---|---|---|---|
| MAI-Image-2.7-Preview | Microsoft | Not listed | 1193, No. 2 | None yet | Unpublished | Editing-only flagship, unbuyable |
| MAI-Image-2.6 / 2.6-Flash | Microsoft | Not listed | 1151 / 1144 | Paid on Foundry | About $0.0389 / $0.0195 per image | Only buyable MAI image models |
| Nano Banana 2.1 Edit | 1151, No. 1 | Tied 1151 | Gemini API | About $0.0336 at 1K | 14 reference images plus masks | |
| GPT Image 2 | OpenAI | 1136, No. 2 | 1136 (high), No. 6 | OpenAI API | About $0.053 at 1024x1024 medium | Most tiers, widest price range |
| Meta Muse Image Edit | Meta | 1129, No. 3 | Not listed | Meta Model API | $0.01 per image | Cheapest volume price |
| Jimeng / Seedream 5.0 | ByteDance | Seedream V4.5 Edit 1043, No. 14 | Not listed | BytePlus / Volcano Engine | About $0.035-0.075 (third-party calibers differ) | Layer separation, unique here |
| Tongyi Wanxiang | Alibaba | Wan v2.7 1042, tied No. 15 | Not listed | Alibaba Cloud | 0.14 RMB Beijing / $0.02007 international | Most generous free quota |
| Kling | Kuaishou | Kling O1 Image 944, No. 26 | Not listed | Kling API | $0.014-0.056 per image | Low board, cheap entry |
Note: price cells differ in resolution tier and billing unit; calibers follow. On the AA v2.1 board, MAI-Image-2.6's 1151 ties Nano Banana 2.1 and Krea 3 (Ultra).
Tool by tool
1. Microsoft MAI-Image-2.7-Preview: strikingly strong, simply unbuyable
By Microsoft's official description, 2.7 targets the edits hardest to get right: whole-image restyling, relighting, rewriting in-image text and symbols while keeping them integrated, restoration, and complex changes requiring object-light-environment understanding. Compared with 2.6, it improves on all 10 test cases and all 7 edit-action categories of Artificial Analysis's evaluation; on category slices of the same preference data, reasoning edits, scene and style, and text-and-symbols all reach the category frontier, with scene and style ahead of the next-best model by 24 Elo. Not seven independent experiments -- seven facets of one voting dataset; do not split them apart.
Keep the version line straight. MAI-Image-2.5 is the previous editing flagship (1127, fourth on ModelGrep); 2.6 and 2.6-Flash are the only MAI image models callable through a paid API; 2.7 is the new editing-only flagship. The 2.6 envelope: up to 5 reference images per request, maximum 1.5K output (1536x1536, shortest side 768px), PNG, both text-to-image and editing; Microsoft says 2.6-Flash generates 2.8x faster than GPT-Image-2-Medium. 2.7's reference-image cap and resolution limits are unpublished; text-to-image is "coming soon" with no date. OrcaRouter's read: generation stays on 2.6, editing migrates to 2.7 -- Flash for volume, 2.7 for the hard jobs.
2. Google Nano Banana 2.1 Edit: the reference-image king, small text still blurry
Generally available October 6, 2026 on Gemini 3.6 Flash, one model for generation and editing. Three editing pillars: conversational multi-turn editing (translate an infographic while everything else stays put), mask-based editing (circle a region, change only that), and up to 14 reference images per request (10 objects plus 4 characters). Resolutions 1K/2K/4K across 14 aspect ratios; thinking levels (Minimal, default Medium, High), search grounding, and a Batch API.
Consistency carries official self-reported numbers: multi-character editing 1106 (previous generation 978), mask editing 1049, multi-reference editing 1066, infographic factuality up from 0.179 to 0.521. Weaknesses in its own model card: at 1K small text still blurs, long passages stay weak, people may not match the reference, left and right occasionally swap. Pricing per Google's page: image output $30 per million tokens, about $0.0336 per image at 1K, $0.0504 at 2K, $0.113 at 4K; Batch API halves output; input tokens rose to $1.50 per million. No free API tier; every image carries an invisible SynthID watermark. Nano Banana 2's deprecation date appears in no official document.
3. OpenAI GPT Image 2: a regular on all three boards, widest price range
1136 and second on ModelGrep, 1136 as "GPT Image 2 (high)" and sixth on AA v2.1, 1461 and third on Arena -- steady everywhere, while its stronger sibling GPT Image 2.5 Sunburst (max) holds the AA summit at 1199. Billing is token-based: image input $8, cached input $2, output $30 per million; text input $5; Batch API halves both directions, and every reference image in an edit adds input tokens. Per the official guide, per-image cost runs $0.006 (low), $0.053 (medium), $0.211 (high) at 1024x1024, and $0.041 (medium) to $0.165 (high) at 1024x1536 -- low/medium/high are quality tiers, not resolutions; the same image can swing 35x across tiers. Speed has a third-party figure: median latency about 4.9 seconds at 1024x1024 (unofficial, third-party measured).
4. Meta Muse Image Edit: the one-cent volume price
Launched July 7, 2026 in Meta AI, the first externally facing image model from Meta Superintelligence Labs -- free everyday, subscription for heavy use; the Meta Model API opened August 25 at $0.01 per image, lowest in the field. It is a reasoning editor: look, plan the layout, then paint; instruction edits land only in the specified region; anchored composition pins series consistency with a few reference images; its chain of thought self-corrects spontaneously. The API mirrors OpenAI's Images API: point an OpenAI SDK at another base_url, model name muse-image-1.0. Official cautions: output caps at 1600px; in-image text varies run to run, so keep it short and rerun when unhappy; multi-image composition may rearrange the layout and is non-deterministic. Consumer tricks: circle and annotate to edit -- remove passersby, restore old photos, change a hairstyle -- with session context carrying multiple turns.
5. ByteDance Jimeng / Seedream 5.0: layer separation, found nowhere else
Region editing offers tap, lasso, box, and scribble pointing; multi-reference composition takes up to 14 reference images for character and product consistency; in-image text renders in 14 languages; deep thinking and realtime web search precede painting; output reaches 2K (a 4K tier on Pro). The signature move is layer separation: a generated poster splits into 10-plus transparent PNG layers carried straight into a design tool -- embedding editing into the design workflow, invisible to every leaderboard. Pricing has two coexisting third-party calibers (one reports Pro $0.075 and Lite $0.035; the other $0.035-0.045 standard, about $0.09 high resolution), so this review quotes the $0.035-0.075 range and defers to the official page. Entrances: the Jimeng app on points membership with daily login credits (Image 4.0's 4K free until November 6, 2026), and BytePlus ModelArk or Volcano Engine for developers, plus a dedicated editing model, SeedEdit 3.0. Speed from the previous generation: Seedream 4.0 did 2K in about 1.8 seconds (third-party). On ModelGrep it is Seedream V4.5 Edit that ranks, 1043 at No. 14.
6. Alibaba Tongyi Wanxiang: the most generous free quota
The dedicated editing endpoint is Wanxiang 2.1 general image editing (wanx2.1-imageedit), covering outpainting, watermark removal, style transfer, instruction editing, inpainting, and restoration by text command. Pricing: 0.14 RMB per image in the Beijing region with 500 free images; $0.02007 internationally with no free quota. Rate limits of 2 requests per second and 2 concurrent tasks rule out heavy lifting. In the newer Wan series, wan2.7-image costs 0.2 RMB and wan2.7-image-pro 0.5 RMB per image; on ModelGrep both Wan v2.7 and v2.7 pro sit at 1042, tied at No. 15.
7. Kuaishou Kling: low board, cheap entrance
The image API family spans Kling Image 3.0, 3.0 Omni, O1, and 2.1, with the official rate card at $0.014-0.056 per image depending on model, mode, and resolution, image outpainting at $0.028, and AI Multi-Shot at $0.07 per run; the Basic membership tier is free with 660 credits monthly, paid tiers 3,000 to 26,000. A third-party platform adds another entry at 0.2 RMB per image (0.4 RMB at 4K). The weakness: Kling O1 Image sits at 944, No. 26 on ModelGrep, bottom among the three Chinese entries and a tier behind the top three in editing; the win is the cheap entrance.
The ledger at ten thousand images
Suppose 10,000 standard images at sticker prices: Muse about $100; MAI 2.6-Flash about $195; Wanxiang 1,400 RMB in Beijing (about $201 internationally); Kling $140 to $560; Seedream 5.0 $350-750 within its quoted range; Nano Banana 2.1 about $336 at 1K, $504 at 2K, $1,130 at 4K; MAI 2.6 about $389; GPT Image 2 from $60 (low) to $2,110 (high) at 1024x1024 -- tier choice alone decides the bill. The free-quota route: Wanxiang's 500 images plus Kling's 660 monthly credits cover small-team experiments; Muse's everyday-free tier suits consumers; Nano Banana 2.1 has no free API tier.
Speed has three citable figures: officially, 2.6-Flash generates 2.8x faster than GPT-Image-2-Medium; third-party medians, about 2.1 seconds for 2.6-Flash, 4.6 for 2.6, 4.9 for GPT-Image-2; and 1.8 seconds for Seedream 4.0 at 2K (both third-party). Nobody else publishes a comparable caliber -- no speed podium.
Cold read
First, a high board score does not mean the instruction was executed. The Arena board's own disclaimer says it is a preference vote that never verifies instruction-following; testing your most common instruction class on real images beats reading Elo. Second, in-image text gives everyone a soft spot: 2.7 makes text-and-symbols a category frontier and stresses scene integration; Nano Banana 2.1 admits 1K small text blurs; Muse varies run to run; Seedream 5.0 sells 14-language text; the MAI family has excelled at English text since 2.0 while Chinese garbles. Third, the consistency watershed is reference count plus masking: 14 references at two tools (Nano Banana 2.1, Seedream 5.0), 2.6 caps at 5, 2.7 unpublished. Fourth, human retouchers are not obsolete: boards measure whether an edit looks like what you asked, not whether it makes business sense -- layer separation, invisible on boards, is worth real money.
FAQ
Q1: MAI-Image-2.7 ranks second -- why can't I use it right now?
A1: Still a Preview, try-able only in the MAI Playground, Arena, and Artificial Analysis; API, pricing, and commercial availability are unpublished, weights closed. Payable Microsoft image models today: MAI-Image-2.6 and 2.6-Flash only.
Q2: Why can't Elo scores from the three leaderboards be compared?
A2: Different scales. The ModelGrep top is about 1151, the AA v2.1 top 1199, the Arena top 1514 -- three independent evaluations. MAI-Image-2.5's 1127 on ModelGrep and 2.7's 1193 on the AA board belong to different boards; a summed "66-point gain" is invalid.
Q3: Which tool edits only a region without re-rendering the whole image?
A3: Three routes: Nano Banana 2.1's masks change only the circled area; Seedream 5.0 offers tap, lasso, box, and scribble; Muse does circle-and-annotate on the consumer side. MAI 2.6 mentions no masking; 2.7's details are unpublished.
Q4: How can a small team with a near-zero budget get started?
A4: Wanxiang 2.1 editing: 500 free images in Beijing; Kling Basic: 660 credits monthly; Muse: free for everyday use; Jimeng app: daily login credits (Image 4.0's 4K free until November 6, 2026). Stacked, a few hundred images a month costs almost nothing.
Q5: Whose in-image text editing is most reliable?
A5: Nobody deserves a blank check. Seedream 5.0 leads with 14-language text rendering; the MAI family excels at English text while Chinese garbles (2.0-era caliber); Nano Banana 2.1 admits small text blurs at 1K; Muse suggests short text and reruns. Test with your own originals before committing.
Related reading