Before August 20, when I used GPT-5.6 Sol Thinking in ordinary Chat on the web client as a ChatGPT Plus subscriber, a single task could run for over 102 minutes.
However, from August 20 until now, every Chat session is forcibly interrupted after only 25–26 minutes, cutting off the task that is still in progress. This has persisted for a full week.
Work that could originally be completed in one go now has to be split into multiple runs. But splitting it up appears to make GPT’s memory problems worse — it more severely forgets the plans and work content from earlier stages, which heavily impacts the continuity of long-running tasks.
This matches an independently preserved case almost exactly. I have a historical ordinary-Chat GPT-5.6 Thinking/Extended turn at 102m26s, while current genuine GPT-5.6 ordinary-Chat turns repeatedly cluster around 25–26m with unfinished work.
OpenAI Support Case #13832863 contains the native historical/current request and turn IDs. The evidence thread is here:
If you are willing, please report your own example through normal OpenAI Support. If they give you a numeric Case Number, posting only that number here or in the thread above would let OpenAI correlate two independently affected accounts. Please keep request IDs, account details, HARs and private chat contents inside Support.