Codex is becoming unusable: extreme slowness and 'at capacity' errors across all models #53591

Summary

Codex has recently become extremely slow and unreliable. Model selection no longer offers a meaningful workaround: after switching models, the application can still fail with the exact same capacity error.

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

Problems observed

• Severe latency: Responses and normal agent progress can take many minutes without clear feedback or visible progress.
• Not limited to one model: The slowdown persists even after changing models and reasoning levels.
• Capacity errors across available models: Multiple model selections fail with Selected model is at capacity. Please try a different model.
• Misleading error guidance: Advising users to select another model is not helpful when other models also return the same error.
• Unreliable service: These failures interrupt normal software development and make Codex difficult to use.

Observed on October 13, 2026.

Expected behavior

• Reliable response times and availability for paid Codex users.
• If backend capacity is exhausted, report the incident clearly rather than implying that the selected model alone is at fault.
• Provide useful retry status or an estimated wait time.
• Distinguish backend overload, account usage limits, queuing delays, and client-side failures.
• Show meaningful progress or an actionable timeout rather than leaving requests pending for many minutes.
• Avoid charging usage for requests that fail before generating a useful result.

Questions for the Codex team

1. Are there known service-side capacity or routing problems affecting multiple models?
2. Why does the UI continue recommending model switching when all options can fail?
3. Is the current extreme latency related to overload, queued requests, or repeated automatic retries?
4. Will OpenAI improve service-status visibility and failure diagnostics for Codex users?

Similar public reports: #53559, #53569, #53528.

This report intentionally contains no repository details, task descriptions, external service names, paths, endpoints, session identifiers, or private workflow information.