"Unexpected status 404 Not Found" in Codex: the 3 causes and how to fix them

Codex stopped answering and printed unexpected status 404 Not Found. Before you reinstall anything, read what comes after the 404: in September 2026 that same line showed up with three different causes in the official openai/codex repository, and each one has its own fix.

Which of the three 404s did you get?

The text after 404 Not Found: tells you the cause. Match it against the table and jump to the right section.

Message fragment	Cause	What to do
Unknown error, url: https://chatgpt.com/backend-api/codex/responses	OpenAI backend failure	Check status.openai.com and wait
The model `gpt-5.5` does not exist or you do not have access to it	Model missing from your account or client catalog	Switch models with /model or -m
url: https://chatgpt.com/backend-api/codex/responses/compact	remote_compaction_v2 = false in config.toml	Delete the line and reopen Codex

"404 Not Found: Unknown error": is it on your side?

No. When the message says Unknown error and the URL ends in /codex/responses, the failure is in OpenAI's backend, and nothing you change on your machine will fix it.

That is what happened on September 3, 2026. Issue #42534 collected more than 30 reports within minutes, from Windows, macOS and Linux, going through Cloudflare edges in different regions (VNO, NRT, SIN, MIA). One report showed that even the model list was failing:

failed to refresh available models: unexpected status 404 Not Found: Unknown error,
url: https://chatgpt.com/backend-api/codex/models?client_version=0.149.1

OpenAI logged the incident as "Elevated errors across ChatGPT and Codex", with Codex Web, Codex API, CLI and the VS Code extension affected, and marked it resolved the same day (status.openai.com).

What to do:

1. Open status.openai.com. If there is an open incident for Codex, stop there.
2. Starting a new conversation or switching models does not help: people who tried it in #42534 kept getting the same 404.
3. Do not rush to reinstall or update. One report in the same issue upgraded the CLI from 0.152.1 to 0.153.0 and the error stayed the same, because the problem was not in the client.

"The model does not exist or you do not have access to it": how do you switch models?

This 404 means the configured model is not available to your account in that client. Switching to a listed model fixes it right away; waiting for the catalog update fixes it for good.

Two real cases from September 2026:

• gpt-5.5 (issues #44477 and #43415, from Sep 7): the model was still available in the ChatGPT app, but the Codex route returned 404. In the same environment, gpt-5.6-sol worked.
• gpt-6-sol (issue #47412, Sep 23): on Windows, Codex Desktop used a bundled CLI (0.153.4) whose catalog had neither gpt-6-sol nor gpt-6-luna, while the standalone CLI 0.156.1 had both.

What to do: inside the session, open the picker and choose a model from the list:

/model

From the terminal, force the model at launch (the -m flag, or --model, lives in codex-rs/utils/cli/src/shared_options.rs):

codex -m gpt-5.6-sol

To make it stick, set it in ~/.codex/config.toml:

model = "gpt-5.6-sol"

Want to see which models your binary knows before picking one? The debug subcommand prints the catalog as JSON:

codex debug models            # catalog refreshed from the server
codex debug models --bundled  # only the catalog shipped in the binary

If the model shows up on the web but not here, the client is the problem, as with Codex Desktop on Windows. Use the newest standalone CLI until a Desktop build ships with the updated catalog.

URL ending in "/responses/compact": why does the conversation die?

If the 404 shows up right after "Context automatically compacted" and the URL ends in /codex/responses/compact, an old flag in your config.toml is to blame. Delete the line and compaction comes back.

Issue #42468 (Sep 3, 2026) documents it: on Codex Desktop 26.901 with core 0.153.0, anyone with this in the file

[features]
remote_compaction_v2 = false

ended up sending compaction to a retired endpoint. The worst part is the cascade: the conversation stays above the compaction threshold, so every following turn tries to compact again and fails the same way. The thread becomes unusable.

What to do:

1. Check the flag states in your binary:

   codex features list

2. Open ~/.codex/config.toml and delete the line remote_compaction_v2 = false. According to the issue itself, setting it to true or removing the line restores compaction.
3. Update. In the current openai/codex source this flag is marked Stage::Removed in codex-rs/features/src/lib.rs, meaning it no longer exists as an option.

None of the three? How do you diagnose the rest?

Run the local diagnostic. codex doctor checks installation, config, login, network and the backend handshake, without trying to repair anything:

codex doctor            # full report
codex doctor --summary  # grouped summary only
codex doctor --json     # redacted version, to attach to an issue

During the Sep 3 incident, one user showed in #42534 how the 404 appears in the report when network, TLS and login are all fine:

handshake transport error http 404 Not Found
ChatGPT inference URL: ... reachable (HTTP 404)

Green network and TLS with a 404 handshake points to OpenAI's side, and you are back to the first section. If doctor flags a proxy, certificate or login, the problem is local and the report already tells you which.

When filing an issue, include the cf-ray and request id at the end of the error message. That is how the reports in #42534 proved it was a backend failure and not a single machine.