Pricing
ChatGPT "Error in message stream": what it means and what to try
Codex or ChatGPT shows "Error in message stream" mid-task? Where users hit it, how it differs from stream disconnected, and the fixes issues describe.
7 min read
Error strings quoted from openai/codex GitHub issues #49954, #54729 and #53648, plus the OpenAI status page (all verified Nov 8, 2026). Every source is listed at the bottom. No affiliate links on this page (how we work).
The exact error: Error in message stream
First, the honest part. You searched “codex error in message stream,” and the reports that actually contain this string were filed in the openai/codex GitHub repo, but they are about ChatGPT web Chat (chatgpt.com), not the Codex CLI or the Codex desktop app. Both issues say so in writing.
From #54729 (verified Nov 8, 2026), in English as displayed:
Error in message streamThat issue states: “This is about ChatGPT Web’s GPT-6 Pro long-running, multi-agent execution, not the Codex CLI or Codex desktop app.”
From #49954 (verified Nov 8, 2026), the UI was in Korean. The reporter quoted the Korean string and gave their own English translation:
메시지 스트림에 오류 발생
("Error in message stream")So the English in #49954 is the reporter’s translation, not a screenshot of the English UI. That issue lists its surface as “chatgpt.com → Chat, not Work or the Codex CLI.”
The one-line answer: in the reports on file, Error in message stream is a ChatGPT web Chat banner that appeared when a long or tool-heavy answer stopped before finishing. No issue gives a root cause or an official fix as of Nov 8, 2026. None of the fetched issues says the Codex CLI or Codex app shows this string, so we are not going to claim it does.
If you are in the Codex CLI or Codex app and your error says stream disconnected before completion, you want Codex stream disconnected. If it says you hit a limit, go to Codex usage limit reached.
Where it shows up, per the issues
ChatGPT web Chat, web research plus the GitHub plugin (#49954)
Opened Oct 1, 2026 and still open on Nov 8, 2026, titled “[ChatGPT Web] Intermittent message stream errors with web research + GitHub plugin since Oct 1.” The reporter says “the model was GPT-5.6 Sol and the reasoning level was High,” and that the error appeared “when a single request combines web research with investigation through the connected GitHub plugin.” (#49954, verified Nov 8, 2026)
The numbers: “1 failure in 3 runs,” with the failing run logged as “Message stream error; no final answer | 55.3–56.4 seconds.” The reporter also noted: “In an earlier failed web chat, navigating away and returning removed the visible error banner while leaving only the incomplete response and tool activity.”
The issue is careful about what it isn’t: “This is a useful intermittent reproduction report, not a root-cause determination.” No comments, no maintainer reply as of Nov 8, 2026.
ChatGPT web, long multi-agent GPT-6 Pro turns (#54729)
Opened Oct 16, 2026, updated Oct 28, 2026, still open. ChatGPT Pro plan, model gpt-6-pro, Chrome on Windows. The turn ran “~29m40s after the prompt” and then: “The page displayed “Error in message stream”. After reloading, the error banner disappeared, but execution did not resume and the answer remained unfinished.” (#54729, verified Nov 8, 2026)
After the reload, the reporter saw stream_status return {"status":"COMPLETE"} while a subagent still reported "status": "inProgress". Parent says done, child says still working. Selina Meyer couldn’t have staged a more Veep moment: the press release goes out reading COMPLETE while half the staff is still in the hallway mid-sentence.
The reporter is strict about the timing, too: “The time alignment is suggestive, not proof of a 30-minute hard limit,” and “I am not claiming to know the precise server-side root cause.” No maintainer reply as of Nov 8, 2026.
What about the Codex Windows app? (#53648)
This is the one people link next to the others, so let’s be clear. #53648 (opened Oct 13, 2026, still open, verified Nov 8, 2026) is a Codex Windows desktop app report on gpt-6.1-sol with reasoning effort ultra, where three turns failed after 9 to 18 minutes. It does not contain Error in message stream. Its errors are:
stream disconnected before completion: An error occurred while processing your request. You can retry your request, or contact us through our help center at help.openai.com if the error persists. Please include the request ID [redacted] in your message.Selected model is at capacity. Please try a different model.A commenter in that issue linked an OpenAI status incident and wrote that it “was marked resolved on October 13.” We did not open that incident page, so we are reporting the comment only. The reporter wrote that the observations “do not establish whether the cause is service capacity, response transport, the client, context handling, local networking, or multiple independent failures.”
Houston, we have a problem, and in all three issues Houston has not written back yet. That’s the part that stings: three detailed, careful bug reports, two of them spelling out exactly which surface they’re on, and no maintainer reply in any of them as of Nov 8, 2026.
What to try (only what the issues and status page support)
There is no official fix on record. Here’s what the sources do support, labeled by where it comes from.
- Check status.openai.com. As of about 8:10 AM IST on Nov 8, 2026, it listed no incident about message streams, Codex or ChatGPT chat streaming. The only active banner was “Elevated error rates with some U.S. business Mastercards,” which is unrelated. So there is no outage to blame on that date.
- Run it again. #49954 logged “1 failure in 3 runs,” meaning the other two runs of the same request finished. That is an observation, not a guarantee.
- Don’t count on a reload to finish the answer. In #54729 the banner disappeared after reloading but “execution did not resume and the answer remained unfinished.” #49954 saw the same thing after navigating away and back.
- If you use the desktop app, note this data point. #49954 says the reporter “tested the same original question in the ChatGPT desktop app → Chat and obtained completed answers on both Plus and Pro accounts,” and calls it “not a fully controlled cross-client experiment.” Treat it as one data point.
- Know the anecdote, but don’t call it a fix. A commenter in #54729 wrote: “If I close the client early after the Pro request has started and wait for several hours then check back, it is much more likely to have successfully completed.” That’s a user’s experience, not guidance from OpenAI.
- Report it with details. Both web issues are useful because they include model, plan, browser, timing and the exact string. Add yours to the matching issue or file a new one with the same details.
If you got stream disconnected before completion in the Codex app or CLI instead, that string itself says you can retry or contact help.openai.com with the request ID (#53648, verified Nov 8, 2026). The Codex stream disconnected post has the full rundown.
Error in message stream vs stream disconnected vs at capacity
No issue explains a technical difference between these strings. The only thing the sources show is where each one appeared.
| Message | Reported in | Source |
|---|---|---|
| Error in message stream | ChatGPT web Chat | #49954, #54729 |
| Stream disconnected | Codex Windows app | #53648 |
| Model at capacity | Codex Windows app | #53648 |
All three issues verified Nov 8, 2026. “Reported in” reflects only the fetched issues.
Don’t read the table as a diagnosis. It’s a map of who saw what where. If your message talks about limits or a reset time, that is a different problem: see Codex usage limit reached.
FAQ
What does Error in message stream mean in Codex?
In the reports on file it is a ChatGPT web Chat banner shown when an answer stopped before finishing; issues #49954 and #54729 give no root cause (verified Nov 8, 2026).
Does the Codex CLI show Error in message stream? None of the fetched issues says so; both reports with this string say they are about chatgpt.com Chat, not the Codex CLI or app.
Does Error in message stream use my Codex limit? Not published; none of the cited issues says whether a failed turn counts against a usage limit.
Is Error in message stream the same as stream disconnected before completion? No issue explains a technical difference; the fetched reports only show the first in ChatGPT web Chat and the second in the Codex Windows app.
Was there an OpenAI outage? As of Nov 8, 2026, status.openai.com showed no matching incident, only an unrelated Mastercard billing banner.
Here’s what I’d do
Check which surface you’re on. If you’re in ChatGPT web Chat on a long, tool-heavy or multi-agent answer, glance at status.openai.com, then run the request again. Don’t expect a reload to finish the old answer, because two reports say it didn’t. If it keeps happening, add your model, plan, browser, timing and exact string to #49954 or #54729.
If you’re in the Codex app or CLI, you most likely have a different string, and Codex stream disconnected is your page. If long Codex turns keep dying on you, our Codex vs Claude Code comparison and AI coding tools compared page are there for a backup plan.
Sources
- openai/codex GitHub issue #49954, “[ChatGPT Web] Intermittent message stream errors with web research + GitHub plugin since Oct 1” (Korean UI string with the reporter’s English translation; open): https://github.com/openai/codex/issues/49954 (verified 2026-11-08)
- openai/codex GitHub issue #54729, “[ChatGPT Web][GPT-6 Pro][Oct 16 update] Long-running multi-agent turn ends around 30 min without final reply” (
Error in message stream; open): https://github.com/openai/codex/issues/54729 (verified 2026-11-08) - openai/codex GitHub issue #53648, “[Windows App / gpt-6.1-sol ultra] Three consecutive turns fail…” (
stream disconnected before completion,Selected model is at capacity; does not contain “Error in message stream”; open): https://github.com/openai/codex/issues/53648 (verified 2026-11-08) - OpenAI status page (no matching incident; only an unrelated Mastercard banner, about 8:10 AM IST): https://status.openai.com (verified 2026-11-08)