The top photo editing model on the Artificial Analysis editing board right now is not from Google. Why am I writing this? Because Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on September 2, 2026, so they are the newest ones in the market, and the perception is that anything with Gemini on it beats most AI models, right. Those two are language models, they write text, they do not edit photos, that job is Nano Banana’s and Nano Banana did not change on September 2. But the perception changed, new Gemini out so Google’s pictures must be best too, and that is what I am here to correct. On the editing board as of 15 September 2026.
Seedream 5.0 Pro from ByteDance is ninth with an Elo of 1100 and Google’s Nano Banana Pro, the paid one, the one on the Google AI Pro plan, is eleventh at 1094.
The ByteDance model is ahead of Google’s premium model and the board prices it at $90 per thousand images against $134. Google’s free one, Nano Banana 2, is eighth at 1101, one single point above Seedream. So the free Google model and the ByteDance model are basically tied, and the Google model you pay extra for is below both.
Google’s own free model and the ByteDance one are neck and neck, and the Google one people pay for is under both of them.
I keep this board open most days because I run product photos through these models for listings and which one I pick is what I pay per image. I’d assumed for a long time that if you wanted the best edit you paid Google for Pro. The board doesn’t say that anymore, and it hasn’t for a while, I just wasn’t reading the price column.
The Editing Board on 15 September 2026, Top 14
| Rank | Model | Maker | Elo | Votes | Released | Board API price per 1k images |
|---|---|---|---|---|---|---|
| 1 | GPT Image 2.5 Sunburst (max) | OpenAI | 1166 | 5,592 | Sep 2026 | $210.7 |
| 2 | GPT Image 2.5 Flare (max) | OpenAI | 1143 | 5,784 | Sep 2026 | $210.7 |
| 3 | MAI-Image-2.6 | Microsoft AI | 1126 | 14,236 | Aug 2026 | $38.9 |
| 4 | MAI-Image-2.6-Flash | Microsoft AI | 1117 | 10,231 | Sep 2026 | $19.5 |
| 5 | GPT Image 2 (high) | OpenAI | 1114 | 13,551 | Apr 2026 | $211.0 |
| 6 | Muse Image | Meta | 1113 | 10,707 | Jul 2026 | $10.0 |
| 7 | MAI-Image-2.5 | Microsoft AI | 1108 | 12,354 | Jun 2026 | $48.1 |
| 8 | Nano Banana 2 (Gemini 3.1 Flash Image) | 1101 | 13,962 | Feb 2026 | $67.0 | |
| 9 | Seedream 5.0 Pro | ByteDance Seed | 1100 | 12,237 | Jul 2026 | $90.0 |
| 10 | MAI-Image-2.5-Pro | Microsoft AI | 1096 | 13,142 | Jul 2026 | $108.5 |
| 11 | Nano Banana Pro (Gemini 3 Pro Image) | 1094 | 12,929 | Nov 2025 | $134.0 | |
| 12 | GPT Image 1.5 (high) | OpenAI | 1093 | 12,619 | Dec 2025 | $133.0 |
| 13 | Reve 2.1 | Reve | 1089 | 1,039 | Jul 2026 | $200.0 |
| 14 | MAI-Image-2.5-Flash | Microsoft AI | 1077 | 12,117 | Jun 2026 | $20.0 |
Top two are OpenAI’s GPT Image 2.5 pair, both September releases, both about $211 a thousand, and both on around five and a half thousand votes so far, so give those a month too. Then three Microsoft models, Meta’s Muse at ten dollars a thousand which is a story for another day, and then the block this piece is about, rows 8, 9 and 11, the three models a normal person actually has access to, with the small lab sitting under them at 13.
Seedream 5.0 Pro Ties the Free Google Model and Beats the Paid One
Seedream 5.0 Pro came out of ByteDance’s Seed lab in July 2026. Same lab that does Seedance for video. ByteDance is not a small company, nobody’s pretending that, but it is not a trillion dollar company either and it has no phone OS, no search engine, and no default slot on anybody’s home screen. It has to earn the rank.
And it has, on this board. 1100 against Nano Banana 2’s 1101 is a tie, the confidence range on both is plus or minus 9, so the one point means nothing. Against Nano Banana Pro at 1094 it is six points up, still inside the noise honestly, but the direction is the wrong way for a model that costs 50 percent more. Twelve thousand votes on Seedream, thirteen thousand on Pro, this is not a thin sample.
I ran a whey tub through Seedream last week for a listing, plain background swap, and it did the swap without touching the tub. Nano Banana Pro would have done that too but Pro also warmed the whole thing up like it was in a kitchen ad, nobody asked for that. Text on the tub, Seedream got the brand right and mangled the 2.2 kg into something like 2.7, so the label check is not optional, and it pushes the blue about ten percent further than it was. That is the model. Does what you said, a bit too colourful, don’t trust it with digits.
What the Three Models Cost, Google Direct and the Board’s Number
The board lists an API price per thousand images. Those are the numbers Artificial Analysis uses for its cost comparison, and they line up with what Google publishes.
- Nano Banana 2, $67 per thousand, which is Google’s own $0.067 per 1K image on the Gemini API. Cheaper at 0.5K, $0.045, and batch mode halves everything if you can wait. And it is free in the Gemini app, which is why it is the default in most people’s heads.
- Seedream 5.0 Pro, $90 per thousand. Nine cents an image. ByteDance’s own Volcano Engine and the resellers land around there, some a bit under.
- Nano Banana Pro, $134 per thousand at 1K or 2K, and it goes to $0.24 an image at 4K. Consumer side you get it on the Google AI Pro plan.
So per image, Google free model 6.7 cents, ByteDance 9 cents, Google paid model 13.4 cents. The one in the middle is ranked above the one on the right.
Forty listing photos a week, that’s about $3.60 on Seedream and $5.36 on Pro. Nobody is retiring on the difference. But I am paying more for the one ranked lower, and the client whose forty photos they are does not know or care which model did them, so what am I paying the extra for exactly.
Where this gets practical, I didn’t want a ByteDance account and a Google Cloud account and an OpenAI account to compare these. Seedream 5.0 Pro, Nano Banana 2 and Nano Banana Pro all sit under one credit balance on Kimg AI so I could run the same photo through all three from one login and see for myself which one I’d pay for, instead of paying each lab separately to find out. That is an observation about how I test, not a recommendation, use whatever front end you already have.
Reve 2.1 Was Third in August and Is Thirteenth Now, and the Votes Column Explains It
I nearly published a different article. Had the draft done, small lab beats Google, then I looked at the votes column.
Reve AI, Inc. is a creative tooling startup based in Palo Alto. In August 2026 Reve 2.1 went to number third on this exact editing board and a lot of people, me included, started writing “small lab beats Google”. Five weeks later it is thirteenth at 1089, below both Nano Bananas.
Look at the votes column. Reve has 1,039 samples. Everything around it has twelve to fourteen thousand. A model on a thousand votes has a confidence range of plus or minus 15, the board says so right there, so its “rank” is really “somewhere between 8th and 16th”. It went to number one on a few hundred early votes that happened to go its way, and as real votes came in it settled where it actually belongs.
That is not Reve being bad, it is a leaderboard doing what leaderboards do on small samples.
Two things to take from that. One, when you see a headline about a new model topping a board, check the sample count before you check the Elo.
Two, Reve at $200 per thousand images is priced like a top five model and is currently a top fifteen model, so whatever it is good at, and it is good at text and layout, it is not a value pick right now.
Who Should Use Which
- One edit a week on your phone. Gemini app, Nano Banana 2, free, eighth on the board. Done.
- Product photos in volume. Try Seedream 5.0 Pro before you pay for Nano Banana Pro. It is ranked above Pro, it’s a third cheaper, and it is on most front ends. Check the label text on your first batch.
- You already pay for Google AI Pro. Fine, keep using Pro, it is eleventh out of seventy odd, it is not bad. Just know you are not paying for the best edit, you are paying for the Google one.
- You saw the Reve headline. Wait for the votes. If it is still top ten at ten thousand samples, it earned it.
And screenshot the board on the day you read this, because the version I’m looking at is 15 September and by the time you’re reading, GPT Image 2.5 might have a Lite, Microsoft might have a 2.7, and Seedream 5.5 is probably already in someone’s beta. The rankings move. The habit of paying Google for the best one is the only thing that doesn’t.
