I am also experiencing frequent and persistent “Selected model is at capacity” errors in the Codex Desktop application, despite having an active ChatGPT Pro 20x subscription.
The issue significantly disrupts my development workflow, particularly when running long tasks or multiple concurrent sessions.
1. Issue Description
The problem has the following characteristics:
-
All models are affected, including GPT-6.1, 6 and 5.6. Switching models does not resolve the issue.
-
A single Codex conversation typically runs successfully for approximately five minutes before encountering capacity errors.
-
Running multiple conversations concurrently significantly increases the frequency of these errors.
-
The errors are intermittent rather than permanent. Tasks may resume successfully but encounter the same problem again.
-
The ChatGPT web interface operates normally and is noticeably more stable.
-
The Codex Desktop application has already been updated to the latest available version.
-
The issue has persisted despite repeated retries and model switching.
2. Diagnostic Evidence
I investigated the local Codex Desktop SQLite logs at:
~/.codex/logs_2.sqlite
The following records were captured on October 9, 2026 :
18:15:49 | INFO | codex_core::session::turn | HTTP_503
18:15:49 | ERROR | codex_core::tools::router | CAPACITY
18:15:51 | INFO | codex_core::session::turn | CAPACITY
18:16:05 | WARN | codex_app_server_transport::transport::remote_control::websocket | OTHER
18:16:06 | WARN | codex_features | OTHER
These diagnostic labels were derived from keyword searches of the local logs. The HTTP_503 label indicates the presence of “503” in the log entry and has not yet been independently verified as an actual HTTP response status.
Additional capacity-related records were observed repeatedly throughout the same period.
The errors appear across multiple internal Codex components, including:
-
codex_core::session::turn
-
codex_core::session::handlers
-
codex_core::tools::router
This suggests that the issue may involve backend capacity allocation, request admission, or error propagation during ongoing tasks, rather than a single-model availability problem.
3. Steps to Reproduce
-
Launch the latest version of Codex Desktop.
-
Start a coding task requiring multiple iterations of reasoning, file operations, and testing.
-
Allow the task to execute continuously for approximately five minutes.
-
Observe intermittent “Selected model is at capacity” errors.
-
Switch to another model, including GPT-5.6, and observe similar failures.
-
Run two or more Codex conversations concurrently.
-
Observe that capacity errors occur more frequently under concurrent workloads.
4. Requested Investigation
Could your engineering team please investigate the following?
-
Account-level or regional capacity allocation: Is my account being routed to an overloaded or unstable serving pool?
-
Backend request failures: Are repeated model requests during long-running Codex tasks receiving HTTP 503 responses or server-overloaded errors?
-
Concurrent session handling: Is there an undocumented concurrency restriction, admission-control mechanism, or capacity allocation issue affecting Pro 20x accounts?
-
Error classification: Is Codex Desktop incorrectly presenting other backend or transport failures as generic model capacity errors?
-
Retry and recovery: Can transient capacity failures be retried automatically without interrupting the entire coding task?
-
Account-specific investigation: Can you check my account’s backend request and capacity error history around October 9, 2026, at 18:15:49 (UTC+8)?
5. Impact
This issue substantially reduces the usability of Codex Desktop for professional software development.
Long-running programming tasks are repeatedly interrupted, and parallel development workflows become particularly unreliable.
As a ChatGPT Pro 20x subscriber, I would appreciate clarification on whether these failures are related to a temporary infrastructure problem, account-specific routing, or an intentional concurrency limitation.
6. Expected Resolution
I would appreciate:
-
Confirmation of the underlying cause.
-
Clarification on whether my account is affected by a known infrastructure or capacity issue.
-
Any available workaround that preserves long-running tasks and concurrent workflows.
-
Information about an expected fix, if this is a known issue.
I use Codex CLI only btw.