Model behavior: GPT-5.4 still stops early on long-running tasks

GPT-5.2 was the least lazy model. -Codex variants always had this behavior of trying to cut corners and stopping mid-task instead of completing it fully. Based on my usage, GPT-5.4 is an improvement over GPT-5.3-Codex in this regard and feels less "codexy" in a good way, but it still has some of the same undesirable behavior traits, like this one.

For context, the task was to orchestrate subagents while auditing the codebase, with the clear and explicit goal of 200 subagents and a limit of 5 at a time. Instead of working until it achieves the goal, it stops early in most runs — typically around 10/200 — and gives me a progress update instead of continuing. In some runs it gets further, but still stops well short of the goal.

A few thread IDs with the same repeating issue of stopping early. Prompts were mostly the same except for the details of the request. On very similar prompts, GPT-5.2 very rarely stopped early — its behavior in general was much more thorough and deliberate.

019ccc56-9327-7712-ace1-da246ac2964b
019ccc58-3115-78a1-b85f-1fd6e5b669bc
019ccc58-06f0-7b80-9b3a-df1e8c390ff6
019ccc58-6e76-7501-85b2-12966ca26ead