Which image model on GMI Cloud should you use?

I sent the same five prompts to every text-to-image model I could reach on GMI Cloud, timed every request from the server side, priced every request with the estimate endpoint, and scored every output against a fixed rubric. Eighty images across thirteen models, $4.37 at list price. The cheapest model in the set scored about the same as the best. The most expensive one landed in the middle. Twelve of thirteen spelled a two-line bakery sign correctly.

Why I ran this

The image catalog on GMI Cloud has grown past the point where "which one should I use" has an obvious answer. There are five Gemini image models alone. GPT Image 2.5 ships in two flavors with five quality tiers each. Seedream has three generations live at once. The public leaderboards rank these models, but a leaderboard tells you how a model does on other people's prompts, at whatever settings the lab chose, on whatever hardware the lab runs. It does not tell you what you pay on GMI, how long the queue takes, or whether the text on a storefront sign comes out spelled right.

So I measured those three things directly: cost, time, and quality, on one afternoon, through one queue, with one set of prompts.

The Models

The thirteen models, with the parameters I used and the list price the estimate endpoint quoted for exactly those parameters:

The estimate endpoint prices the exact parameters you submit, and that can differ from the catalog card: GPT Image 2 at medium quality quoted $0.0547 against a $0.06 card. Always estimate with the parameters you will send. The Seedream models do not go below 2K, so their prices buy four times the pixels of the 1K models. Luma also returns 2048x2048.

How I measured

Cost is the list price from the estimate endpoint, before any organization discount, for the exact parameters submitted.

Time is measured on the server, not on my laptop. The queue records created_at when it accepts a request. The output lands in Google Cloud Storage and the object carries a Last-Modified header. The difference is the number I report. It includes queue wait, generation and upload, and excludes my polling and network. I ran requests one at a time, so no model competed with another for the queue.

Quality is a 0 to 5 score per prompt from one rater, me, looking at the contact sheets with all thirteen outputs side by side. Two prompts also have a binary check that does not depend on taste: did the sign read exactly "ROSETTA BREAD CO." with "est. 1987" under it, and are there exactly three apples to the left of the bowl. One rater means a gap of a tenth of a point is no gap at all. The arena Elo numbers further down are the vote count.

The five prompts, each chosen to stress one thing:

1. Portrait. A woman in her sixties, silver hair, freckles, window light from the left, 85mm, grey background. Tests skin, lighting, and whether the model drifts to a younger face.

2. Signage. A bakery storefront with a painted sign reading "ROSETTA BREAD CO." and a second line "est. 1987", a bicycle by the door. Tests text rendering.

3. Spatial. Three red apples to the left of one blue bowl, a yellow pencil across the top of the bowl, rain in the window behind. Tests counting, left/right and attribute binding.

4. Illustration. An isometric data center on a floating island, pastel, flat vector, no text, no labels. Tests style control and a negative instruction.

5. Product. Matte black headphones on a lime green backdrop, rim light, reflective floor. Tests studio lighting and materials.

A note on the numbers. Every number in this post comes from the GMI Cloud MaaS inference engine on a Tier 5 account, so your cost and wait times will vary with your account tier, queue load on the day, and the provider's own capacity. If you need faster or more predictable turnaround, a dedicated endpoint runs your model on reserved capacity and cuts the queue wait out of the picture.

The Cost Picture

estimate_generation on GMI Cloud, Oct 14 2026

At the settings I used, one image ranges from $0.013 on GPT Image 2.5 Flare to $0.134 on Gemini 3 Pro Image, a tenfold spread. Seven models sit between $0.035 and $0.055. The Seedream and Luma prices buy 2K output, so per pixel they are among the cheapest in the set.

The quality tiers on the OpenAI models stretch that range further. GPT Image 2.5 Flare goes from $0.006 at low to $0.211 at max for the same 1024x1024 frame, a 35x range inside one model, and every tier spelled the bakery sign correctly. GPT Image 2 at high quality is $0.213. Nano Banana 2.1 is $0.041 at 1K, $0.058 at 2K and $0.121 at 4K. Bria FIBO charges the same $0.04 at 1MP and 4MP. Hy-Image 3.5 at 4096x4096 is $0.032.

The Speed Picture

queue created_at vs storage Last-Modified, 63 runs

The Gemini family is the fastest group by a wide margin. Flash-Lite returned every image in four to eight seconds. Nano Banana 2.1 was steady at 14 to 15 seconds. The slow end is GPT Image 2 at medium, Seedream 5.0 Pro at 2K, and Luma.

If you call Luma from a synchronous client, give it two minutes: one of its five images took 86 seconds, and it was worth the wait.

Test 1: Portrait

Three models produced a portrait I would hand to an editor without a second pass: GPT Image 2, GPT Image 2.5 Flare, and Nano Banana 2.1. All three kept the freckles, kept the age, kept the grey background, and put the light where the prompt asked. Nano Banana 2.1 had the most natural skin of the set. Seedream 5.0 Pro was a close fourth.

Gemini 3 Pro Image took a different reading of the prompt and made an environmental portrait, the subject looking out of a window with the frame in shot. It is a good photograph, but it is not the close-up editorial portrait the prompt asked for. Gemini 3.1 Flash-Lite had the one face in the set with plastic-looking skin. Luma UNI 1.1 cropped tight and flattened the lighting. Hy-Image 3.5 went dark and dramatic, side-lit with bare shoulders, which reads as a different brief.

Test 2: Signage

This is the one I expected to separate the field, and it did not. Twelve of thirteen models rendered "ROSETTA BREAD CO." exactly, in capitals, on a sign above the door. All twelve also got "est. 1987" under it, though Gemini 2.5 Flash Image made the second line almost too small to read and Luma made it faint.

The exception is Hy-Image 3.5. It printed the name inside the window, stacked over three lines, with no "est." line at all. It is the only model in the set that did not follow the placement.

Two models added text I did not ask for. Gemini 3.1 Flash-Lite put "FRESH BREAD DAILY" on the window and invented a brand name on the headphones in test 5. Seedream 5.0 Lite also chose its own aspect ratio here, returning 2848x1600 for a size of "2K", where its sibling returned a square.

The GPT Image 2.5 Flare quality sweep ran on this prompt. All five tiers, from low at $0.006 to max at $0.211, spelled both lines correctly. The tiers differ in the bread texture and the chalkboard lettering, not in whether the sign is right.

Test 3: Spatial

Eleven of twelve models that returned this prompt put three apples to the left of the bowl with the pencil across the rim and rain on the glass. Gemini 3 Pro Image shows two apples clearly and perhaps a third hidden behind them. Hy-Image 3.5 got the objects right but the window behind is dark and the rain is not visible.

I also ran Nano Banana 2.1 at all three thinking levels on this prompt. Minimal, medium and high all produced the correct arrangement. On a prompt this size the extra reasoning tokens bought nothing I could see. The same held for Gemini 3.1 Flash-Lite with thinking set to high.

Seedream 5.0 Lite is scored on the two prompts it ran, portrait and signage.

Test 4: Illustration

Every model produced an isometric island with racks and a turbine, and none added text, so the negative instruction held across the board. The differences were in how "flat vector" was read. GPT Image 2.5 Flare, Nano Banana 2.1, Bria FIBO, Seedream 4.0 and Gemini 3 Pro Image gave the cleanest flat renders. GPT Image 2 drifted toward a shaded 3D look. Hy-Image 3.5 drew a domed building with a pipe and a pond, which is an island but not a data center. Luma made the sparsest scene.

I ran Seedream 5.0 Pro twice on this prompt. The two islands are different and equally correct, which is a useful data point on variance: the model does not collapse to one composition.

Test 5: Product

This was the most even test. Gemini 3.1 Flash Image, Nano Banana 2.1, GPT Image 2, GPT Image 2.5 Flare, Seedream 4.0 and Seedream 5.0 Pro all delivered a usable advertising frame with the rim light and the reflection. Gemini 3 Pro Image cropped so tight that the headband leaves the frame. Gemini 2.5 Flash Image drew a white glow around the product that reads as a cut-out. Hy-Image 3.5 made the headphones wired, with a cable, on a darker green. Hy-Image at 4096x4096 was better than its own 1024 output on this prompt, and costs $0.032 instead of $0.024.

Price against quality

Mean rubric score across the five prompts, against list price:

There is no line through this cloud. The cheapest model scored about the same as the top of the set, and from a single rater a tenth of a point is not a ranking. The most expensive one is ninth. The middle of the price range holds everything from 3.9 to 4.7. If I had to pick on these five prompts alone, GPT Image 2.5 Flare at medium is the value pick, Nano Banana 2.1 is the pick when fourteen seconds matters more than four cents, and GPT Image 2 is what I would reach for when the brief is a portrait.

What the parameters did

I ran fifteen extra requests to see which knobs move the output.

• Quality tiers (GPT Image 2.5 Flare). Low, medium, high, xhigh and max all passed the sign test. Low is $0.006. If your prompt is text-heavy and your canvas is 1024, start at low and only climb when you can point at a defect.

• Resolution (Nano Banana 2.1). 1K, 2K and 4K kept the same composition and the same correct sign. The API returns two URLs for every request. At 1K the two files are pixel-identical. At 2K and 4K the first URL is a 1024x1024 preview and the second is the full-resolution file. Read the second one. One image is generated and billed, not two.

• Thinking (Nano Banana 2.1, Gemini 3.1 Flash-Lite). No visible change on the spatial prompt at any level.

• Style flag (Bria FIBO). style_id photoreal gave a visibly better portrait than the default: better skin, better light. It is free. Use it for photographic prompts.

• 4MP (Bria FIBO). Same $0.04, 2048x2048 instead of 1024x1024.

• 4K (Hy-Image 3.5). $0.032 instead of $0.024, and the product shot was its best output.

What the public leaderboards say

Artificial Analysis runs a blind pairwise arena. Their AA-Image-T2I v2.1 board, dated September 2026, and my five-prompt mean side by side:

LMArena's text-to-image board, as mirrored on sophon.at three days before I ran this, has the same top three: GPT Image 2.5 Sunburst, GPT Image 2.5 Flare, GPT Image 2.

The arena order and my order agree at the top and at the bottom. GPT Image 2.5, GPT Image 2 and Nano Banana 2.1 lead both. Hy-Image 3.5 sits near the bottom of both. The disagreements are the useful part. Bria FIBO scores 849 on the arena and 4.5 with me, because my prompts reward exact text and object placement, which FIBO handles well, and the arena rewards whatever voters prefer across thousands of prompts. Gemini 3 Pro Image is seventh on the arena and ninth with me, mostly because it reinterpreted two of my five prompts.

Note the settings. The arena ranks the OpenAI models at high or max quality. I ran them at medium, which costs a quarter as much, and they still led. You do not need the max tier to get the arena-leading behavior on ordinary prompts.

Key Takeaways

1. Price does not predict quality. The $0.013 model scored about the same as the top of the set and the $0.134 model finished ninth.

2. Text rendering is solved at the top of the catalog. Twelve of thirteen models spelled a two-line sign correctly. The OpenAI low tier at $0.006 did too.

3. Gemini is the speed family. Flash-Lite at 4 seconds, 2.5 Flash at 6, 3.1 Flash at 10, Nano Banana 2.1 at 14. Nothing else came in under 11 seconds.

4. Start at the lowest quality tier and climb only on a defect. On GPT Image 2.5 Flare the tiers cost 35x apart and produced the same sign.

5. Thinking levels did not move the output. Nano Banana's three levels made no difference on a spatial prompt.

6. Nano Banana returns two URLs per request. They are the same image, with the second one at full resolution for 2K and 4K.

7. Estimate with the exact parameters. The estimate endpoint prices what you send, and that is the number you pay.

Quick Reference

Try It on GMI Cloud

Every request in this post went through the GMI request queue. Submit, then poll:

curl -s https://console.gmicloud.ai/api/v1/ie/requestqueue/apikey/requests \
  -H "Authorization: Bearer $GMI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-image-2.5-flare-generate",
       "payload": {"prompt": "Street photograph of a small bakery storefront. The painted wooden sign above the door reads \"ROSETTA BREAD CO.\" in serif capitals, with a smaller line \"est. 1987\" underneath. Morning light, a bicycle leaning by the door",
                   "size": "1024x1024", "quality": "low", "output_format": "png"}}'

curl -s https://console.gmicloud.ai/api/v1/ie/requestqueue/apikey/requests/$REQUEST_ID \
  -H "Authorization: Bearer $GMI_API_KEY"

The Question to Answer

Which of the five prompts looks like your workload? If it is signage and product shots, the cheapest tier of the cheapest OpenAI model is enough and you should spend the savings on more variations. If it is portraits, pay the four extra cents for GPT Image 2 or Nano Banana 2.1. If it is throughput, the Gemini Flash models return before most others have left the queue.

Ready to build? Get Started