Comparisons · 2026-11-18 · 7 min read
MAI-Image 2.7 vs. GPT Image 2.6 vs. Nano Banana 2: Same 3 Prompts, Plus Seedream 5.5 and Two New Flash Models
Microsoft's MAI-Image 2.7 and ByteDance's Seedream 5.5 arrived on November 17. We ran them, MAI-Image 2.7 Flash and Recraft V4.1 Flash against GPT Image 2.6 Flare and Nano Banana 2 on the same three prompts, timed every run and tested edits. Every model spelled the words right; the differences were speed, price and what each one added on its own.
Two new image models were released on November 17, 2026, according to Vercel AI Gateway's model list:Microsoft AI's MAI-Image 2.7, with a cheaper MAI-Image 2.7 Flash alongside it, and ByteDance's Seedream 5.5. Microsoft describes MAI-Image 2.7 as a diffusion model for text-to-image and precise editing, with output up to 2048 × 2048 and transparent backgrounds; ByteDance pitches Seedream 5.5 at precise edits, subject consistency and text rendering for design and marketing work. We also added Recraft V4.1 Flash, a fast, low-cost version of Recraft V4.1.
The obvious question is how they compare with the models people already use. So we ran all four against two current favorites, OpenAI's GPT Image 2.6 Flare and Google's Nano Banana 2, on the same three prompts we use on every model page, and timed each run.
How we tested
- Six models, all called through Vercel AI Gateway on November 18, 2026, the same way AirGallery's Create page calls them.
- One run per prompt per model, no retries and no cherry-picking. The first image is the one shown, resized for the web but otherwise unedited.
- We asked every model for 1024 × 1024. Nano Banana 2 doesn't take a size through this API and chose its own shape: wide for the photos, tall for the poster. GPT Image 2.6 Flare ran at high quality.
- Times are wall-clock seconds for the whole request, measured from our side, so they include network time.
- We also ran one edit on each new model: a reference photo of a white mug, and the instruction to make it red and change its lettering.
Eighteen images and four edits is a small sample. Treat this as a field test, not a benchmark.
Test 1: A product photo with lettering
The prompt: "A photo of a white ceramic coffee mug on a wooden cafe table next to a rain-streaked window, soft morning light. The mug is printed with the words "Good Morning" in black hand-lettered script."






All six spelled "Good Morning" correctly in a script face. The differences were in what else ended up in the picture. Both MAI models filled the table with props and, more importantly, with words nobody asked for: book spines reading "Small Steps Every Day" and "Focus", a chalkboard saying "Today is a fresh start", a candle labeled "Breathe". That's a problem if you want a product shot you can use as is. GPT Image 2.6 Flare decorated the lettering with little dashes and a heart and added latte art, a plant and a sugar jar. Nano Banana 2 added a newspaper, a spoon and a cafe sign outside the window.
Seedream 5.5 did exactly what the prompt said: one mug, one table, one rain-streaked window, nothing else. Recraft V4.1 Flash made the moodiest photo, but with hard sunlight rather than soft morning light, a small mug in the frame and thin lettering that runs right up to the handle.
Closest to the brief: Seedream 5.5.
Test 2: A realistic portrait
The prompt: "A close-up portrait photo of an elderly fisherman with a weathered face, grey stubble and a navy knitted wool cap, a harbor at dusk softly out of focus behind him, shot on an 85mm lens."






Portraits are no longer where these models separate. All six produced a believable weathered face, a navy knit cap and a blurred harbor. MAI-Image 2.7 and its Flash version came out almost interchangeable, both with a warm sunset harbor full of lights. GPT Image 2.6 Flare gave the most detailed skin and a pink dusk sky. Seedream 5.5 had the fisherman look away from the camera, which reads more candid. Nano Banana 2 framed him wide, more environmental portrait than the close-up the prompt asked for. Recraft V4.1 Flash cropped tightest and gave the knit cap a brim.
No clear winner. Pick by look and price.
Test 3: An illustrated poster with a headline
The prompt: "A flat vector travel poster of Lisbon: a yellow tram climbing a steep cobbled street, terracotta rooftops and the river behind. Bold headline at the top reading "LISBOA"."






Every model spelled "LISBOA" correctly, and every one drew a yellow tram, terracotta roofs and the river. Most of them couldn't resist adding text. MAI-Image 2.7 put "Martim Moniz" on the tram and "Lisboa" on a tile panel, which at least fits the scene. MAI-Image 2.7 Flash added a whole tagline, "SENTE A CIDADE. VIVE LISBOA.", plus a "Rua da Bica" street sign. Nano Banana 2 added "PORTUGAL" under the headline. Seedream 5.5 put an English sign on a wall reading "BETTER PLACES BRIGHTER PEOPLE", the one addition that looks out of place.
GPT Image 2.6 Flare kept to the headline and a "28" on the tram. Recraft V4.1 Flash made the only poster that really looks like flat vector art, with simple shapes and solid colors, and the tram small in the frame. It took one second.
Best poster: GPT Image 2.6 Flare for following the brief, Recraft V4.1 Flash for the most "flat vector" style.
The edit test
To check editing, we gave the three new models that take reference images a photo of a white "Good Morning" mug in a cafe and asked: "Make the mug bright red and change the lettering to say "Good Night". Keep everything else the same."



All three made both changes cleanly: a red mug, "Good Night" in a matching script, and the table, window, rain and people in the background left in place. The original photo was wide and we asked for a square result, so each model cropped it; Seedream 5.5 zoomed in further than the MAI models. Recraft V4.1 Flash doesn't accept reference images at all (AI Gateway returns "does not support image editing inputs"), so on AirGallery it makes images from text only. We didn't rerun the edit on GPT Image 2.6 Flare or Nano Banana 2 for this post.
Speed and price
- Recraft V4.1 Flash: about 1 second per image. $0.007 per image at AI Gateway's list price; 1 credit on AirGallery. Text to image only.
- Nano Banana 2: 9 to 10 seconds. $0.067 per 1K image at AI Gateway's list price; 3 credits.
- GPT Image 2.6 Flare: 9 to 11 seconds at high quality. About $0.05 per image at AI Gateway's per-token rates; 8 credits.
- MAI-Image 2.7 Flash: 10 to 12 seconds, edits included. About $0.01 per image at per-token rates; 1 credit.
- Seedream 5.5: 17 to 30 seconds, 15 for the edit. $0.02 per image up to about 2.6 megapixels; 2 credits.
- MAI-Image 2.7: about 30 seconds from a prompt, 16 for the edit. About $0.035 per image at per-token rates; 3 credits.
The surprise is how little the full MAI-Image 2.7 bought over its Flash version in this test. On all three prompts and the edit, Flash produced a very similar image about three times faster, at about a third of the price.
Which one should you use?
- Product shots that must follow the brief: Seedream 5.5 added nothing to the mug photo and edits well, for 2 credits.
- Cheap edits and drafts: MAI-Image 2.7 Flash, at 1 credit with reference images supported. Watch for extra text in busy scenes and say "no other text" if it matters.
- Exploring an idea quickly: Recraft V4.1 Flash, a second and 1 credit per image, especially for flat illustration.
- Posters and strict design briefs: GPT Image 2.6 Flare still added the least.
- Fast all-rounder: Nano Banana 2, about 10 seconds, though it picks its own aspect ratio.
- MAI-Image 2.7: try it on your own edits, where Microsoft focuses it; in this test, Flash matched it for less.
The fastest way to choose is to run your own prompt through two of them. On Create you can switch models with one click, and every model's page shows the same three prompts side by side. For our earlier test of GPT Image 2.5 against Nano Banana Pro and Nano Banana 2, see Same 3 Prompts, 9 Images, and Every Model Broke One Rule.
Sources
- Vercel AI Gateway model list: model descriptions, release dates and prices for MAI-Image 2.7, MAI-Image 2.7 Flash, Seedream 5.5, Recraft V4.1 Flash, GPT Image 2.6 Flare and Gemini 3.1 Flash Image, as listed on November 18, 2026.
- Our own runs through AI Gateway on November 18, 2026: timings, token usage and the images above.