How to Fix ChatGPT "Selected model is at capacity" (Rate Limits and Degraded Answers)
When ChatGPT shows Selected model is at capacity, the current model is usually being rate-limited. Another common symptom is a clear drop in answer quality, which the community calls "degraded answers" (降智).
The error is reported by many users worldwide; it shows up across the OpenAI Community forum, GitHub repos and X. What can be confirmed is that this is account risk control, but OpenAI does not publish the rules or explain how to recover. Below is a statement from an OpenAI employee on limiting model access:
The gist:
"Even while a paid subscription is active, access to certain models or features may be temporarily limited based on account activity. We can't provide more details about these checks. The system re-evaluates access automatically; once the activity causing the limit stops, access returns to normal."
After ChatGPT or Codex is rate-limited or degraded, the full recovery order is: stop using the account, move to a clean environment, use it lightly for a while, and only then add heavier usage back.
Some users shared a simpler approach with a lower time cost; try it first:
- Reset the password
- Set up keys
- Add MFA
- Sign out signed-in devices
- Cool down for 12 hours before using the account again
If the simple approach does not work, follow the full process below.
Step 1: Stop the usage that can trigger risk control again
Sign out of everything: web, app, desktop and Codex.
Turn off agent subagents, do not enable Ultra, and do not attach sub2api, reverse proxies, multi-account or multi-session setups.
Step 2: Let the account sit for at least 24 hours
Do not keep signing in to check. If it is still degraded after 24 hours, keep it idle; do not run heavy tests the moment the time is up.
Step 3: Move to a clean home broadband or residential IP
Do not use datacenter, server or shared IPs. Best case, one fixed egress per account. Some users find ordinary home broadband is not enough and switch to a cleaner residential line.
Step 4: Clean up the account itself
Change the password, enable 2FA, and confirm nobody else is using the account.
Step 5: Test lightly through official entry points only, adding things back in order
- Sign in to ChatGPT in the browser first, using Work or 6 Pro, with English prompts, a few messages, the way you would normally use it.
- Once the web client looks normal, try ChatGPT Desktop, still in English.
- Once desktop is normal, test Codex mode.
- If all of that is normal, consider reconnecting the account to tools like cpa or sub2api the next day; keep one account per egress and lower concurrency first. Reverse proxies are not recommended: OpenAI has dedicated Cyber abuse rules against reverse-proxy use, and repeated warnings lead to bans.
Step 6: How to use the account during recovery
For the first few days, ask in English. You can let it answer in Chinese, but do not hold long Chinese conversations right away. Do not jump straight into many sessions, high concurrency or back-to-back long tasks. If a test still comes back degraded, sign out, keep the account idle, and do not force it.
Step 7: After it looks recovered, watch it for about 3 days
Keep sessions few for those 3 days, keep subagents off, and do not enable Ultra or max out usage right away. Some users report normal use after this step; others degrade again after a couple of days.
The whole flow in one line: sign out and idle the account, switch to clean home broadband, change the password and enable 2FA, use official clients in English with low concurrency for a while, then add tools and concurrency back gradually.