ChatGPT Plus 5-hour window: why the OAuth pool returns 503

What the rolling 5-hour window actually is

ChatGPT Plus accounts are subject to a rolling usage window that OpenAI resets on roughly a 5-hour cadence. The exact published budget is not part of OpenAI's documented error surface for API clients — there is no row for 504 or for "rolling window exhausted" on the OpenAI platform's Error codes guide. What you observe from the relay side is a moving quota that drops to zero on a single Plus account once sustained traffic passes through it.

The window is per-account, not per-model. Switching from one model to another on the same Plus OAuth login does not reset the timer, and it does not give you a fresh budget.

Why your client gets 503, not 429

The relay does not surface OpenAI's internal quota signal as a 429 to your client. The chain we observed in production on Aug 7, 2026 looked like this:

# abbreviated relay log (actual lines from journalctl -u sub2api)
WARN  openai_chat_completions.upstream_failover_switching
      upstream_status=401  account_id=16  switch_count=1  max_switches=10
      model=gpt-5.6-sol    stream=true
WARN  openai_chat_completions.account_select_failed
      error="no available accounts"  excluded_account_count=0
INFO  http request completed  status_code=503  latency_ms=16  model=gpt-5.6-sol

Step by step:

1. The relay picks a ChatGPT Plus account (account_id 16 in our logs) and forwards the request.
2. OpenAI Plus returns 401 to the relay, meaning the rolling window is exhausted and the account's token is effectively revoked for inference traffic.
3. The relay marks the account exhausted and tries failover — up to 10 switches in our observed config.
4. Every other Plus account in the pool is in the same shape, so each switch returns 401 too.
5. After the switch budget runs out, the relay returns 503 with a JSON body that mentions "no available accounts".

The relay absorbs the OpenAI Plus quota signal and translates it into its own 503. Your client never sees the upstream 401 directly.

Why this is different from a 401 to your client

A simple way to confirm before debugging your own config: /v1/models with the same key returns 200 with a JSON model list. If that route passes and /v1/chat/completions returns 503, the failure is in the inference account pool, not in your request shape.

Read the relay log the right way

Three log lines together pinpoint the ChatGPT Plus 5-hour window failure. Look for them in order:

If excluded_account_count is greater than zero, failover actually tried every account and each was excluded. If it is zero, the relay never even started a second selection cycle. Both numbers map to the same root cause: the ChatGPT Plus pool cannot serve the request right now.

What to do as a client

1. Stop the tight retry loop. max_retries=0 for the duration of the diagnostic.
2. Drop concurrency by at least half. The window is rate-based, not request-count-based.
3. Wait for the rolling window. Recovery is gradual — a 09:00 burn gives partial recovery at 14:00, not full recovery.
4. If your workload is steady and you cannot tolerate the cycle, switch to a different tier: an API key from a paid OpenAI API plan does not share the Plus 5-hour window.
5. Capture request_id, UTC time, model, and the relay's "no available accounts" body before contacting support. Redact the API key.

What an operator should check

If you operate the relay, the same production pattern gives you a smaller set of real actions:

• Confirm the failing requests all share account_id on the Plus OAuth pool, not the API-key pool.
• Check whether the Antigravity token refresher is logging Account N needs refresh on schedule. If refreshes are healthy and upstream still returns 401, the issue is upstream quota, not token caching.
• Track upstream_failover_switching count per day. A sustained count (we saw 241 in 7 days, roughly stable per day) is the early-warning signal that the Plus pool is at its 5-hour ceiling.
• Tell the user to reduce concurrency; do not raise the relay's proxy_read_timeout in response to this failure. The timer is irrelevant when the upstream already returned 401.

FAQ

Related guides