I run a small tool that turns a song into a cover you can upload to a distributor. Every cover starts as an image-model render, so the model choice is the biggest line in the cost sheet and the biggest lever on how the product looks.

Six days after launch I asked the obvious question: the models I picked are from earlier this year — should I be on the newest ones? Seedream 5.0 had shipped. GPT Image 2.5 had shipped. Grok Imagine Image 2.0 had shipped.

So I ran the newest ones against what I already use, with the same prompts, on the same two API platforms, and measured four things: what it costs, what size actually comes back, how long it takes, and whether the picture matches the brief. This post is the numbers, the pictures, and the decision — including the part where I got burned.

The constraint that matters more than quality

Before any of the numbers make sense you need one fact about the market. Apple Music asks for cover art of at least 4000 × 4000 pixels. TuneCore and CD Baby, two of the three big distributors, accept nothing above 3000 × 3000. No single file satisfies both, so a cover tool has to be able to produce a genuine 4096 px image, natively or through an upscaler.

That one fact eliminates half the candidates before quality is even discussed. Any model whose top output is 2K or 3K costs you a second call to an upscaler (Topaz on fal.ai is $0.08 per image), which triples or quadruples the price of the render.

The incumbent, and how it got the job

The current stack is Seedream 4.5 for paid renders and Z-Image for free ones, both through kie.ai; Nano Banana 2 for restyling a user's own photo and Topaz for upscaling, both through fal.ai.

Seedream 4.5 won the first benchmark on September 5 against Flux 2 Pro, GPT Image 2 and Nano Banana Pro — 36 images, \(1.84 total, one render per cell, no cherry-picking. Two things decided it: all four models produced covers that read as photographs at thumbnail size, so quality was not the separator; and Seedream was the only one that could return a 4096 px image in one call (its high-quality setting, same \)0.0325 as 2K) in 22 to 33 seconds. GPT Image 2 took 72 seconds median. Nano Banana Pro cost $0.09 and once took 172 seconds.

Round two: the same prompts, the new models

I reused three prompts from a real production run — the abstract, scene and portrait directions for a lo-fi single called Midnight Vending Machine — so every model got identical input, and I already had Seedream 4.5's output to put next to them. One call per cell. No retries, no picking the best of N.

Two of those rows are the reason to always measure instead of reading the spec sheet.

GPT Image 2.5 with resolution: "4K" returned 2880 × 2880. Not 4096. That is below Apple Music's minimum, so the "4K" label on the request is not the same thing as a 4K file in the response. On fal.ai the same model tops out at 1024 for a square. It is also the best-looking output of the round — the vending machine has products in it, the bus window has rain on it, the prompt is followed more literally than by anything else — and twice as fast as GPT Image 2 was in September. None of that matters if the file cannot be uploaded.

Seedream 5.0 Lite is cheaper than 4.5 and does return 4096 — but it took twice as long, and on the abstract prompt it dropped the subject. The brief asked for the brushed-steel vent panel of a vending machine; 4.5 drew exactly that, 5.0 Lite drew a generic scratched metal surface. One image proves nothing about the model's average, but it does prove it is not a free upgrade.

Grok Imagine Image 2.0 I did not even render: 2K maximum on both platforms means an upscale on every image, and the September benchmark already showed the incumbent does not lose on realism.

Left to right: Seedream 4.5, Seedream 5.0 Lite, GPT Image 2.5. Top to bottom: abstract, scene, portrait.

Where the newest model did win

The photo-restyling step was a different story. The product can take a user's own photo and re-shoot it in one of six looks — black-and-white film, two-colour Riso, 70s Kodachrome, instant film, direct flash, warm grain. Five of them worked on the old model. "Direct flash" did not, on night photos: six prompt variants over two sessions, and the model would only darken the frame instead of adding a hard on-camera flash. I had written it off as a model limitation.

Nano Banana 2 (edit endpoint, fal.ai, 1K, 23 seconds) did it on the first try with the unchanged production prompt: hard frontal light on the face and jacket, background dropped to black, a hard shadow on the chain-link fence behind the subject, identity and pose untouched.

Source photo · the best of six attempts on the previous model · Nano Banana 2 with the same prompt.

It costs $0.08 per image against $0.02 before. For a feature that charges four credits a look, that is still a small fraction of revenue, so it shipped the same day. I then ran all six looks once more on the new model to make sure the other five still held. Five did. The Riso prompt said "printed on off-white paper", and the new model took that literally and drew a paper border around the picture — which is exactly what a distributor's reviewer flags as a mockup. One sentence in the prompt fixed it.

The decision

• Paid renders stay on Seedream 4.5. It is the only model in the table that is cheap, fast and 4096 in one call. Newer did not mean better for this job.

• Photo restyling moved to Nano Banana 2. It solved a real, documented failure, at a cost the feature's price absorbs.

• Nothing moved to GPT Image 2.5 or Grok. Not for quality reasons. For pixel reasons.

Three things I would do differently

Read creditsConsumed, not the price list. kie.ai's task record tells you what a call actually cost after the fact. GPT Image 2.5 "4K" was 16 credits ($0.08); Seedream 5.0 Lite "ultra" was 5.5 ($0.0275). Neither number was on the model page.

Download with a browser user agent. My first bench run lost a completed render because the result CDN returned 403 to Python's default urllib agent. The task was billed, the file was gone, and I had not logged the task id. Now every cell writes its record to disk before the download starts.

Pick the platform on stability, then let price break ties. Both platforms carry the same models. One of them hung an upscale queue for a day last week and returned 500 on two text-to-speech models the day of this test. The rule I wrote down: the stable platform by default, the other only when it is clearly cheaper — which, for Seedream 4.5 at 4096, it is.

The tool is Album Art Creator. The pre-flight check that enforces the 4000-vs-3000 problem is open source as cover-art-check on npm. Prices and sizes above were read from the two platforms' model pages and task records on 14 September 2026.