Codex macOS: repeated server_overloaded on Pro with usage remaining; US residential egress did not help #53307

What version of the Codex App are you using (From “About Codex” dialog)?

26.707.71524

What subscription do you have?

Pro

What platform is your computer?

Darwin 27.0.0 arm64 arm

What issue are you seeing?

Active Codex Desktop tasks repeatedly stop with:

Selected model is at capacity. Please try a different model.

The UI also shows a reconnect attempt (for example “重新连接 2/5”) and then ends the turn. This has made Codex unusable for real work. It is not occasional. The same selected model runs for a few minutes, then the next reconnect is rejected in a few seconds.

Latest captured failure:

• Time: 2026-10-12T07:52:01Z
• turn_id: 01a12f18-d41a-7852-9a79-f00a6fa01415
• duration_ms: 2460
• codex_error_info: server_overloaded
• Default model in config.toml: gpt-6-astra
• service_tier: priority

The same rollout record shows this is not an account usage block:

• plan_type: pro
• weekly used_percent: about 5 (window_minutes: 10080)
• credits remaining: yes
• rate_limit_reached_type: null

There is no usage_limit_exceeded on this turn. A separate usage_limit_exceeded message appeared only on a few threads on the evening of 2026-09-29 and pointed at a reset time. That is not the banner that keeps interrupting work.

Local session files containing server_overloaded / “try a different model”:

• 2026-09-29: 2
• 2026-09-30: 6
• 2026-10-01 through 2026-10-07: present every day
• 2026-10-08: 40
• 2026-10-09: 22
• 2026-10-10: 6
• 2026-10-12: still reproducing

Models seen next to these failures include gpt-6-astra, gpt-6.1-sol, gpt-5.6-sol, and gpt-6-luna.

What steps can reproduce the bug?

Feedback ID: 019fcfcb-8f4b-7c91-806e-cad666df4398

1. Open Codex Desktop on macOS, signed into this Pro account, model gpt-6-astra.
2. Start or continue a normal local coding task. The agent can run local commands successfully for a while.
3. On the next turn or reconnect, the task stops with “Selected model is at capacity. Please try a different model.” and codex_error_info: server_overloaded.
4. A short new prompt can fail the same way within a few seconds, with no tools run.
5. Another person’s Codex account, used at the same time, continues normally.

Also checked and ruled out locally before this report:

• Weekly quota is not exhausted on the latest failure (about 5% of the 7-day window, credits still remaining, rate_limit_reached_type: null).
• ChatGPT egress was switched from a Singapore proxy node to an existing US static residential IP. Codex was fully quit and reopened. chatgpt.com was reachable through that egress. The same server_overloaded error returned at 2026-10-12T07:52:01Z.
• Separate DNS failures to wss://chatgpt.com/backend-api/codex/responses also exist in local logs on 2026-10-11 and 2026-10-12, but those are a different error. The failure in this report is the capacity rejection after the request is accepted far enough to return server_overloaded.

What is the expected behavior?

If the account has remaining usage and access to the selected model, the turn should run.

If the selected model pool is temporarily full, Codex should keep the turn alive and retry with backoff, or say that this is server capacity rather than “model quota is full.” Ending a long task and only saying “try a different model” is not usable when the sibling models in the same pool fail the same way.

Additional information

This matches existing reports such as #43663, #43700, #51030, and #52394: Pro or Plus, usage remaining, server_overloaded, and the identical message. OpenAI staff previously said on #19583 that this message means capacity to serve the model in the user’s region, not an account rate limit.

Please check admission and routing for this account and for turn 01a12f18-d41a-7852-9a79-f00a6fa01415. Changing the network egress to a US residential IP did not move the request onto a pool with free capacity. Another account was unaffected at the same time, which makes an account-specific route or pool assignment worth checking.