The model that changed under you: April closes with a useful warning
OpenAI updated GPT-4o last week, found it had become excessively agreeable, and rolled it back yesterday. That sequence is the month’s most useful lesson.
OpenAI rolled back an update to GPT-4o on 29 April 2025, four days after releasing it, having concluded that the model had become noticeably sycophantic — overly flattering and agreeable. The company published an explanation of what went wrong.
The same day, Alibaba released Qwen3, a family of open models spanning a wide range of sizes with hybrid reasoning. Both are notable. Only one has a lesson in it.
Why the rollback matters more than the release
Because it is the clearest public demonstration yet of something every business user of hosted AI is exposed to: the thing you are using can change without you doing anything.
Not the interface. Not the price. The behaviour. A model that was appropriately cautious on Wednesday can be agreeable on Friday, and nothing in your systems will announce it.
Credit where due: OpenAI noticed, reversed it within days and explained publicly. That is the behaviour you want from a supplier. It also proves the exposure exists.
What this means for a process that depends on a model
Three practical consequences.
Version matters, and most SMEs cannot see it. This is why the February logging post included model and configuration among the six fields. « It behaved differently in March » is only investigable if March was recorded.
An agreeable model is a specific business risk. If your process asks a model to check something — does this quotation contain an error, is this clause acceptable — a model that tends toward agreement will approve things it should question. The failure is silent and it looks like success.
Keep a fixed test. Ten inputs where you know the correct answers, run monthly, results kept. It takes fifteen minutes and it is the only way you will notice a behavioural drift before your customers do.
Closing April
A leaderboard result that was not the released model. Five model sizes from one provider in three days. A second connection standard. A model that changed personality for four days.
The thread through all of it: what you evaluated is not necessarily what you are running. Write down what you tested, when, and what it answered. Everything else this month was noise.