I am consistently seeing long-running ChatGPT tasks stop after approximately 23–26 minutes when using GPT-5.6 Sol High on a Plus account.
I have reproduced this both with workflows based on OpenAI Secure MCP Tunnel and with separate coding tasks performed through Bash scripts. Because the same behavior occurs without relying on my MCP gateway, it does not appear to be specific to my gateway implementation or to Secure MCP Tunnel itself.
Observed behavior:
- The task continues normally for approximately 23–26 minutes.
- ChatGPT then stops the run with:
ChatGPT, you ran out of time to interact with tools and it is time to send a message back to the user. - The same approximate 23–26 minute limit is reproducible in OpenAI Secure MCP Tunnel workflows and in Bash-based coding tasks.
- When Secure MCP Tunnel is involved, the gateway itself does not crash.
- Other concurrent agents using the same gateway continue operating normally.
- The same behavior has continued after previously identified gateway-side async issues were fixed.
Expected behavior:
Previously, comparable long-running tasks could continue for well over 100 minutes. I am looking for a practical way to keep long-running tasks working beyond the current approximately 23–26 minute termination point.
Environment: - ChatGPT Plus
- GPT-5.6 Sol High
- ChatGPT web
- Developer Mode when using Secure MCP Tunnel
- OpenAI Secure MCP Tunnel for MCP-based reproductions
- Separate Bash-based coding tasks reproduce the same approximate 23–26 minute termination behavior
- Observed repeatedly in September 2026, JST (UTC+9)
I asked OpenAI Support for help with this issue, but Support does not appear to be able to resolve it. Therefore, please understand that simply suggesting that I contact Support would not be useful, as I have already done so.
I would like to know: - Is anyone else consistently seeing a similar 23–26 minute limit?
- Does it also occur with GPT-5.6 Sol High for you?
- Has anyone identified what triggers it: elapsed time, tool turns, call count, tokens, or something else?
- Has anyone found a reliable workaround that allows genuinely long-running tasks to continue beyond this limit?
I am especially interested in reproducible observations from users who have actually measured the duration rather than occasional timeout reports.