A GPT Image edit fix: why explicit settings need regression checks

The correction

On January 9, 2026, OpenAI fixed GPT Image 1.5 and chatgpt-image-latest edits that used high fidelity despite an explicit low setting.

Test intent as well as HTTP success

A request can complete successfully while behaving differently from the setting your application selected. Keep a few saved edit requests that exercise important options. Review their output and cost when a provider announces a behavior correction.

For this kind of change, compare a source with fine detail against a simpler source. Keep the prompt and output settings constant, and record the explicitly selected fidelity value in your own diagnostic data.

Avoid mistaking a correction for an outage

After a fix, output characteristics or cost can differ from the previous period. Check whether your expectations were based on documented behavior or on the earlier defect. Reverting application code may not restore a provider-side behavior.

If an option is important enough to expose in your user interface, explain what it is intended to control and maintain a corresponding acceptance check.

Improve the support trail

When investigating an unexpected edit, keep a redacted request, time, model ID, selected settings, and request identifier. Avoid placing API keys or confidential source images in a public support ticket. Good evidence shortens the path from a vague visual complaint to a reproducible case.