ChatGPT Stuck on Thinking: Triage for Teams

When ChatGPT sits on Thinking or loading, start a new chat and compare model, browser and network before feeding more into a conversation that has already stalled.

Symptoms

The reply indicator spins, "Thinking" stays on screen, or the answer starts and then freezes mid-sentence. Sometimes the page itself never finishes loading and the sidebar is empty. Some users see it only with one model, some only in one very long conversation, and some see it across the whole office at the same time. Note which of those describes you before you do anything else.

Likely causes, most common first

1. The model is genuinely working. Reasoning models and deep research tasks can take several minutes by design. A spinner for two minutes on a complex request is not a fault. A spinner for twenty minutes on "summarise this paragraph" is.
2. A service incident or degraded performance. Slow or hanging responses are the usual shape of an OpenAI capacity problem. The status page will often show "degraded performance" rather than an outage.
3. The conversation is too long. Each reply carries the whole history. Very long chats, especially with several large files, slow down and eventually hang. This is the most common cause for one user with one bad conversation.
4. Streaming is being broken by the network. ChatGPT streams replies as they are generated. Proxies and TLS inspection that buffer the response, or filters that block WebSocket connections, make the reply look stuck until it fails. Works on a hotspot, fails in the office, is the tell.
5. Browser tab or extension trouble. A background tab throttled by the browser, a memory-starved machine with forty tabs open, or an extension that intercepts requests.
6. An out-of-date desktop or mobile app. The apps are updated often. Old versions can hang after a server-side change.
7. Switching networks mid-reply. Moving from wifi to mobile data, or a VPN reconnecting, drops the stream. The reply looks frozen but the connection is gone.

Step-by-step fix

1. Wait a sensible time for the model you chose. For a reasoning model on a hard question, give it a few minutes. For the default model on a simple prompt, thirty seconds is plenty.
2. Check the OpenAI status page. If it shows degraded performance, wait or switch to a lighter model for now.
3. Stop the reply and send it again once. If the second attempt hangs too, do not keep retrying in the same conversation.
4. Open a new chat and ask something trivial. If that answers instantly, the old conversation is the problem. Ask the old one for a short summary of what matters, copy it into the new chat, and carry on there.
5. Switch model. If the reasoning model hangs and the default model answers, it is a model-side capacity issue and not your setup.
6. Reload the page, or fully quit and reopen the desktop or mobile app. Update the app if an update is waiting.
7. Try a private window with extensions off, then a phone hotspot with the VPN off. Whichever change fixes it names the cause.
8. If a large file is attached, remove it and try the question without it. Then attach a smaller extract.

What IT should check

• Proxy or firewall logs for buffered or terminated long-lived connections to chatgpt.com during the reported times. A filter that inspects and re-assembles responses will break streaming.
• Whether WebSocket connections to OpenAI domains are allowed, and whether a recent policy change added category blocks or TLS inspection for AI services.
• Endpoint health on the affected machines: memory pressure, browser version and the extensions installed by policy.
• Whether the issue follows one user, one device or one network segment.

The ChatGPT WebSocket network error note covers the streaming and firewall side in more detail. If the underlying problem is a conversation that has grown too large, the same habits described in the context window too long note apply to ChatGPT.

Habits that stop it recurring

Keep one conversation per task and start a new one when the task changes. Put reference documents into a Project rather than re-attaching them to every chat. Use the default model for routine work and reserve reasoning models for questions that need them. Give staff a one-line rule: if it hangs twice, new chat, then hotspot, then IT.

When to escalate to OpenAI

Raise a ticket when a fresh chat with a short prompt hangs in a private window on a second network, with no incident on the status page. Include the model, the time in UTC, the app or browser version, whether the account is on a Business or Enterprise workspace, and the shortest prompt that reproduces it. Do not send the stalled conversation itself if it contains client material.

Source basis

• OpenAI: service status
• Microsoft: web content filtering in Defender for Endpoint
• Google Chrome: clear cache and cookies