Image generation resolution control in Codex / ChatGPT app / web ~broken since mid-June

Environment

  • Surface: CLI / IDE extension / App / Web / Cloud
  • Codex version: any
  • Model and plan or API-key authentication: ChatGPT subscription

Bug

Since mid-June there seems to be a set pixel/token budget allocated to any image generation not done with the paid API. Meaning an image you prompt in the browser, in Codex CLI, in the ChatGPT app and so on will always have a resolution that amounts to about 1.5 million pixels.
If you prompt for 16:9 4K, or 16:9 1080p, you always get something like 1672×941. If you want a 128x128 image, you still get 1254x1254 instead.

It’s understandable that OpenAI might want a set cost per image for subscriptions, but surely this can’t be at the peril of users being unable to control how big their images actually are?
Even if a 1920x1088 or 3840x2160 image is not permitted for budget or plan separation reasons, surely users should be able to get a 256x256 image generated so they can make, dunno, icons?

This GitHub post has a basic chart showing it affects all resolutions:
openai/codex/issues/28723#issuecomment-4730508791
(direct link not permitted)

In my own tests, I found it does not matter if the relevant endpoints are called with a size parameter and/or the requested resolution is named in the prompt. At most you can control the aspect ratio, but never the actual resolution budget.

Reproduction

Steps or a minimal prompt/example:
Prompt for a 3840x2160 16:9 image in any supported interface. Observe you get 1672×940, 1672×941, 1673×940, or similar.

Diagnostics

When it happened (with timezone) and how often:
Always.

Relatedly, it does not seem possible to call the gpt-image-2.5-flare and gpt-image-2.5-sunburst models through subscriber quota endpoints either.
In Codex CLI, gpt-image-2 is hardcoded:
openai/codex/blob/main/codex-rs/ext/image-generation/src/tool.rs#L65