# Ran 30 rounds of cube edits on Seedream 5.5, GPT Image 2.5 and 2.6, is 2.6 really a downgrade? — retained record

> Retained reddit record from reddit:singularity.

Canonical: https://highsignal.app/data/records/c296f9a090c46233

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:singularity)  published 2026-11-19 09:20 ingested 2026-11-20 02:37

[original source ↗](https://www.reddit.com/r/singularity/comments/1y0541i/ran_30_rounds_of_cube_edits_on_seedream_55_gpt/)

## retained text



```
I wanted to try Seedream 5.5 and GPT Image 2.6 since they came out around the same time. I'd been hearing people say 2.6 was worse than 2.5, so I brought 2.5 into the comparison too.

I had all three edit the same scrambled cube for 30 rounds, following the same sequence of 90-degree face turns. Each output became the input for the next round. All images were 2048×2048, with both Image models set to high.

I used the Atlac Cloud CLI in Codex to quickly call all three models and run the editing chains.

Generating 30 images cost $1.03 for Seedream, $3.05 for Image 2.5, and $3.02 for Image 2.6. Median generation time per image was 39 seconds, 51 seconds, and 69 seconds, respectively. Seedream was clearly the cheapest and fastest of the three.

Looking through the results, 2.5 gave me the funniest moment. It had been building up all this colorful texture in the background, then suddenly hid it with a clean gray background at round 28. Might have something to do with the “keep it pure gray” instruction I'd been giving it, but it really looked like it found a shortcut there, haha.

2.6 accumulated noticeably less of that colorful background texture than 2.5 over the repeated edits. The difference became pretty obvious as the rounds went on. It was slower, but for this kind of multi-round editing, that looks like a decent improvement to me.
```



document fetched 2026-11-20 00:29 · canonical https://reddit.com/r/singularity/comments/1y0541i/ran_30_rounds_of_cube_edits_on_seedream_55_gpt
