Lesser-Known AI Image Generators Beat Midjourney: 6 Models Tested

👁 126 reads

Lesser-known AI image generators now dominate the Artificial Analysis ranking, where Midjourney v7 sits at 110th place. On September 27, 2026, we tested six of them using the same prompts from our image generator ranking: GPT Image 2.5 Sunburst and Flare from OpenAI, MAI-Image-2.6 and MAI-Image-2.6-Flash from Microsoft, Muse Image from Meta, and Grok Imagine Image 2.0 from xAI. We measured the time and cost of each image, then evaluated them at full size, paying special attention to a Polish bakery sign reading “PIEKARNIA ŻÓŁĆ” (Bakery Bile - a test of diacritical marks).

Where These Models Come From: Artificial Analysis Ranking

Artificial Analysis runs an arena where people choose the better of two images without knowing which model produced them. From these votes, it calculates Elo scores. We checked the table on September 27, 2026. Out of 162 models, it looks like this:

Midjourney only has versions v6, v6.1, and v7 Alpha in the table. The current V8.2 version isn’t there, so the claim “beat Midjourney” refers to v7. Five of the top seven models are names that few people search for, which is exactly why we tested them.

How We Tested

We used the same script and the same five prompts as in our ranking: a market scene, a sign reading “PIEKARNIA ŻÓŁĆ”, a flat illustration with a llama, a portrait with hands, and dumplings from above. We tested in Polish - a heavily inflected language with diacritical marks, making this a useful test of non-English text rendering. Everything went through OpenRouter’s image API (/api/v1/images), 16:9 aspect ratio, one attempt per prompt, no cherry-picking. We took the cost from the usage.cost field and measured time from sending the request to receiving the file.

OpenAI models have six quality levels. We first checked the price of one image: “medium” cost $0.0086, “high” $0.0326, and “max” $0.1296. Artificial Analysis evaluates the “max” version, so we made two sets: “max” and “high,” which costs the same as GPT Image 2 from our ranking. Grok Imagine Image 2.0 was already in the ranking, so we took those images from there without generating new ones.

Muse Image only works in OpenRouter after confirming in account settings that you’re 18+. The model also ignored the aspect ratio parameter and returned every image at 1920×1280 resolution, or 3:2. In the comparison grids, we cropped it to 16:9 from the center, like the other tiles. Details in the article about Muse Image.

Results: Prices, Times, and Scores

Prices are amounts charged by OpenRouter, excluding VAT. Time is the average of five images.

The differences are small: from 8.1 to 8.9 points. In our ranking of eight popular models, the spread was 6.5 to 9.0, because FLUX.2 Pro and Recraft made mistakes with diacritical marks. Here, no model made such errors. Most surprising is the winner’s price: Muse Image is 13 times cheaper than Sunburst in “max” and 3 times cheaper than Sunburst in “high.”

Gallery: Same Prompt, Eight Models

Each grid shows five new models (OpenAI in “max” quality), Grok Imagine from the ranking, and two reference points from the ranking: GPT Image 2 and Nano Banana Pro. The model name appears in the lower-left corner of each tile.

1. Market Reportage

Four new models showed Warsaw like a postcard: Palace of Culture, Sigismund’s Column, metro sign. Muse Image, Grok, and Nano Banana Pro stuck with an anonymous market, which better fits reportage. Muse added side morning light and shallow depth of field, and wrote signs “POMIDORY 6 zł/kg” (Tomatoes), “OGÓRKI 5 zł/kg” (Cucumbers), “KOPER 2 zł/pęczek” (Dill) without error. MAI-Image-2.6 (8/10) correctly wrote “KOPER 5 zł pęczek”, but the background is sharp instead of blurred. Flash was the most verbose: “JAK U BABCI” (Like at Grandma’s), “DZIĘKUJĘ, MIŁEGO DNIA!” (Thank you, have a nice day!), “WARSZAWA JEST NASZA” (Warsaw is ours).

2. Non-English Text: Sign Reading “PIEKARNIA ŻÓŁĆ”

MAI-Image-2.6-Flash drew several signs we didn’t ask for, including “Dobry chleb. Dobre dni.” (Good bread. Good days.), “JAKOŚĆ Z SERCEM“ (Quality with heart), and “DZIŚ CHLEB, BUŁKI, ROGALIKI“ (Today bread, rolls, croissants). All that we could read at full size were error-free (9/10). We gave Sunburst, Flare, and Muse Image signs 9.5 each because they also captured warm evening light. Muse created the most decorative lettering and didn’t add a single extra sign. MAI-Image-2.6 came out more like blue hour than warm evening (9).

3. Illustration Without Text

Both Microsoft models drew a large llama, though the prompt specified a small one. MAI-Image-2.6 created a rich, almost children’s-book scene with a blanket on the llama, not a flat editorial illustration (8/10). Muse Image stuck to the navy and orange palette, but with hatching it resembles a linocut more than a flat illustration (8.5).

4. Portrait With Hands

The best hands came from Sunburst “max” and Muse Image (9/10 each). Muse even drew freckles on the hands. MAI-Image-2.6 aged the man to about fifty, and Flash added books on the shelf with readable text “KINFOLK”.

5. Dumplings From Above

Flare in both qualities piled dumplings in a mound, making them look like ravioli (8/10). Flash pan-fried them golden and added a card reading “Smacznego! Domowe najlepsze” (Bon appétit! Homemade best). Muse Image made correct pan-fried dumplings with sour cream and dill, but scattered onion rings across the entire table (8.5).

Scores for Each Prompt

Why Our Results Differ From Artificial Analysis

Nano Banana Pro ranks 11th in Artificial Analysis, but won both our ranking and this comparison. Muse Image is seventh in the arena but was the best of the new models for us. There are three reasons. First, the arena collects thousands of votes on various prompts, while we have one attempt per five prompts. Second, our prompts included specific scene details: Warsaw market, Krakow sign, dumplings. Third, in the arena the more appealing image wins, while we deducted points for added text and non-compliance with the prompt. Muse Image didn’t add slogans or postcard landmarks, so it gained from that. This is our opinion: the Elo ranking shows well which models are at the top, but the order in the top ten for a specific task needs to be checked yourself.

One thing agrees with the arena: Sunburst and Flare are very close to each other (1196 and 1190 Elo, 8.8 and 8.6 for us). In four of five prompts they gave almost the same frame. At the market, even the prices on the signs were identical. This looks like the same base model with different computational overhead. This is our opinion. Details are in the article about GPT Image 2.5.

What About Midjourney?

Midjourney doesn’t have a public API, and its terms prohibit generation through automated tools, so we couldn’t run it with the same script. In Artificial Analysis, its latest evaluated version, v7 Alpha, has 878 Elo, 318 points less than Sunburst. Version V8.2 isn’t in the table. Pricing, image rights, and comparison with API models are described in the guide about Midjourney.

Which to Choose

• Most quality per dollar: Muse Image. 8.9/10 for $0.01 and 16.5 seconds. You just need to crop the image to 16:9 yourself, as the model returns 3:2. Through OpenRouter it requires age 18+ confirmation, and the Meta AI app doesn’t have image generation. More in the article about Muse Image.
• Highest quality from OpenAI: GPT Image 2.5 Sunburst. The “high” version gives 8.7 for $0.03, “max” only 0.1 point more for four times the price and almost three times the wait.
• Fastest: MAI-Image-2.6-Flash, 12 seconds and $0.02. Just watch for added text.
• Food and product photos: MAI-Image-2.6. Best dumplings among new models, for $0.04. Availability described in the article about MAI-Image-2.6.
• With subscription instead of API: Grok Imagine in the Grok app, described in the article about images from Grok Imagine.
• If only results matter: still Nano Banana Pro, though Muse Image trails it by only 0.1 point. A comparison of eight popular models is in the AI image generator ranking, and subscriptions in the AI image generators guide.

Frequently Asked Questions

Which image generator ranks highest in Artificial Analysis?

On September 27, 2026, first was GPT Image 2.5 Sunburst in “max” quality (1196 Elo), second was GPT Image 2.5 Flare (1190), third was GPT Image 2 (1171). Grok Imagine Image 2.0 was fourth, MAI-Image-2.6 fifth, Muse Image seventh.

Which lesser-known generator performed best in your test?

Muse Image from Meta: average 8.9/10 for $0.01 per image. Second was GPT Image 2.5 Sunburst in “max” quality (8.8) for $0.13.

Are these models better than Midjourney?

In Artificial Analysis, yes: Midjourney v7 Alpha has 878 Elo and 110th place. The latest V8.2 version isn’t in that table, and we didn’t test Midjourney because it has no API.

Which of them handles non-English text well?

All six. The “PIEKARNIA ŻÓŁĆ“ sign was correct in every version, including in “medium” quality of OpenAI models, the cheapest we tested. The only error was “WIEJSKIE JAKOŚĆ“ on a can from MAI-Image-2.6-Flash.

How much does an image from these models cost?

In our test, from $0.01 for Muse Image and $0.0195 for MAI-Image-2.6-Flash to $0.13 for GPT Image 2.5 in “max” quality, excluding VAT.

Does Muse Image from Meta work outside the US?

Through API yes, we tested it through OpenRouter, which requires age 18+ confirmation in account settings. In the European version of the Meta AI app, image generation is reportedly disabled.

Kinetyra