They removed the old model. Users made them bring it back

When GPT-5 launched on 7 August, OpenAI retired the earlier models and routed all users to the new one. The reaction was loud enough that within a few days GPT-4o was restored as an option for paying subscribers.

Reporting framed this as users being attached to a model’s personality. That is part of it. The more useful reading, for a company, is different.

What people were actually complaining about

Work that had been tuned stopped behaving the same way. Prompts people had refined over a year produced different output. Not worse in a benchmark sense — different, which for a process that runs every week is the same problem.

This blog made the point twice before. On 30 April, about a model updated in place. On 18 June, about the difference between a preview and a version under a stability commitment. The August episode is the largest instance so far, and the first where the vendor reversed course.

The lesson is not « avoid new models »

It is that the upgrade you did not choose is the one that costs you.

Two practical consequences. First: for anything your business runs repeatedly, prefer a supplier arrangement where a version is pinned and changes are announced. Second: keep the handful of prompts that matter in a file you own, because if you have to re-tune them, you want the starting point.

The 23 July post on writing reusable prompts is the cheap insurance for exactly this. Three weeks later it got its demonstration.

The encouraging part

The reversal happened within days, because customers said so clearly and in large numbers. That is not nothing. A market where suppliers can be moved by their users is healthier than one where they cannot.

But you should not have to rely on a public outcry to keep your Tuesday-morning process working. Ask about version pinning now, in the quiet weeks, rather than the next time it happens.