Problem Solutions
ChatGPT Work Keeps Returning HTTP 500 Errors — What Users Tried and What Actually Helps
ChatGPT Work has been failing intermittently since October 27, 2026, according to a detailed report on OpenAI’s developer forum. The user, a ChatGPT Plus subscriber, reports that Work suddenly began failing — including in brand-new conversations — with the message “Hmm…something seems to have gone wrong.” Browser developer tools confirm the underlying error: HTTP 500 Internal Server Error on the request that carries the conversation.
The most frustrating part is that the problem comes and goes. Work occasionally recovers and can even complete complex tasks — running a cloud browser task, executing commands, generating downloadable files — then starts failing again minutes later, on a simple question, with nothing changed on the user’s side.
What users are seeing
The symptoms, as reported: the failure message appears frequently; both existing and new Work conversations are affected; the issue persists across models, including GPT-6 Sol High and GPT-6.1 Sol High; and image attachments have a second fault, where Work receives the image filename but reports that the actual image content is unavailable. The forum listing shows the thread is recent and still without a confirmed fix from OpenAI.
What has already been ruled out
This is where the report is genuinely useful, because the user tested almost everything a support script would ask for: Microsoft Edge on Windows including incognito mode, mobile access, different network connections and VPN nodes, different models and thinking settings, new and existing Work conversations, and repeated sign-in and retry attempts. None of it gave a reliable fix.
Two facts narrow the cause. First, both desktop and phone tests used VPN connections, though with different nodes, so a VPN routing issue has not been completely ruled out. Second, and more telling, regular ChatGPT and Codex Work continue to function normally on the same account. When the same account works in two products and fails in a third, the fault most likely sits in that third product’s backend, not in the user’s device, network or account.
How to keep working until it is fixed
There is no confirmed fix yet, so the honest advice is about limiting the damage, not curing the fault. Keep Work tasks short and save every output to a file the moment it arrives, so a failure minutes later costs you nothing. When Work fails, run the same task in regular ChatGPT or Codex Work, which are unaffected on the same account. If a task fails, wait a few minutes and retry rather than hammering it — the intermittent recoveries show the backend sometimes clears on its own. And if you report it, include the developer-tools evidence, the HTTP 500 and the request path, because that is what separates a real backend fault from a vague “it broke” report.
Don’t burn time on what is already ruled out: reinstalling, switching networks, changing models or starting fresh conversations have all been tried and none of them fixed it. This is a different fault from the regular-ChatGPT breaking reported yesterday — streaming interrupts, capacity errors and chats flagged too long — because this one is Work-only, and regular ChatGPT is fine. If OpenAI publishes a cause or a fix, that is the update to watch for; until then, short tasks, saved outputs and the unaffected products are the way through.