OpenAI slashes Codex capacity errors
OpenAI rolled out a series of client and service-tier backoff patches that cut 'model at capacity' errors for Codex users from nearly 7% to 0.085%. The updates resolve WebSocket retry loops and propagate Retry-After delays to prevent cascading rate-limit failures during coding sessions.
Capacity errors during autonomous coding runs are lethal because naive client retries quickly turn a brief service hiccup into a thundering-herd outage. By coordinating client-side backoff and propagating Retry-After signals across WebSockets, OpenAI finally stabilized the infrastructure underpinning long-running agent workflows.
- –Peak capacity error rates hit 6.82% of daily active users in mid-October before falling to 0.085% across six iterative retry-handling releases.
- –The root cause was client-side reconnection storms hammering overloaded endpoints rather than individual user token quotas expiring.
- –Propagating explicit retry delays for NO_MORE_RETRY and maintaining backoff across WebSocket reconnects stops agent loops from crashing mid-refactor.
- –Reliable uninterrupted execution is mandatory as Codex shifts from simple autocompletions to multi-hour autonomous repository migrations.
DISCOVERED
39m ago
2026-11-16
PUBLISHED
54m ago
2026-11-16
RELEVANCE
AUTHOR
thsottiaux
