OpenAI returns 503 / queue / timeout
separate outage, rate limit, and quota
As of 2026-10-20, when OpenAI ChatGPT, Codex, or the API returns 503, 429, or a timeout, the key is not always dead. Open status.openai.com first, then decide whether to wait, change model, or change endpoint. The status page lists components separately: Chat Completions and Codex in ChatGPT Desktop each have their own status, so read the component line and not only the overall banner. A paused ChatGPT subscription and a burned credit window are different problems.
Updated 2026-10-20
One key for Claude, GPT and leading Chinese models, billed per token, up and running in minutes.
- Billed per token — live rates on /models
- Pay by card (Visa / Mastercard / AMEX), Apple Pay, Google Pay or crypto
- Self-service: top up and activate instantly
Four signals that are not the same bug
OpenAI-side capacity or routing
503, 502, and elevated errors on the status page are usually not a bad prompt. Back off. Do not retry in lockstep.
Your rate or concurrency
429 is your account or org window. Lower concurrency before you swap models.
Credits or the billing window
Green status plus a rejected request is more often quota than an outage. That write-up is /gpt-6-astra-quota-burn, not this page.
Read the official status page first
When status.openai.com says fully operational, debug your 429 and quota before you assume a global outage.
How to handle this class of failure
As of 2026-10-20, the official error-code guide names both codes: 503 - Model temporarily overloaded, 429 - Rate limit reached for requests. The first is a temporary capacity problem on the model: the guide says to wait for the Retry-After header if present, otherwise lengthen the delay between retries, and check the status page for an active incident if it continues. The second is your own request rate limit: wait for Retry-After if present, otherwise use exponential backoff with jitter and cap the retries, and send fewer requests or tokens. OpenAI errors, queues, and timeouts recur; they are not tied to one incident id. The order stays fixed: read the current banner on status.openai.com and the latest title on the history page, then check whether the HTTP code is 5xx or 429, then decide to wait, shed load, or point the same key at another model family. ChatGPT being down does not prove the API is down. A Codex error does not prove the Responses API died with it.
Incidents in 2026-09 that we could verify
On 2026-09-03 reporting described a multi-component outage hitting ChatGPT and Codex across at least 15 service areas, with the status page flipping to degraded performance around 14:56 UTC. On 2026-09-04 at 6:30 p.m. Eastern, The New Stack updated that GPT-6 Astra was available on ChatGPT for paying users whose plans include it. On 2026-09-10 the history list includes Elevated errors affecting ChatGPT Work. On 2026-09-11 status.openai.com read We’re fully operational and not aware of any issues affecting our systems; llmlatency.dev the same day says Last verified September 11, 2026.
Timeline
Around the 2026-09-03 GPT-6 Astra launch, ChatGPT and Codex saw a multi-component outage; reporting cites at least 15 service areas and a degraded-performance banner near 14:56 UTC.
On 2026-09-04 at 6:30 p.m. Eastern, public reporting updated that GPT-6 Astra was available on ChatGPT to all paying users with plans that include it. Being locked out of a launch and sitting through an outage are not the same event.
On 2026-09-10 the status history records Elevated errors affecting ChatGPT Work. On 2026-09-11 the live page copy is fully operational. The incident ended; the diagnostic order did not.
Confirmed versus do not write as fact
Checkable on the status page or in reporting (as of 2026-10-20)
As of 2026-10-20: status.openai.com reports All Systems Operational overall; the status page gives each component its own status, and both Chat Completions and Codex in ChatGPT Desktop are operational. The official error-code guide writes 503 as a temporary model overload and 429 as a reached request rate limit. The official record for 2026-10-14 is Degraded performance in Codex and ChatGPT Work; 2026-10-19 has File Download errors; 2026-10-15 has Widespread feature degradation; 2026-10-13 has a degraded-performance entry for GPT-6.1 Sol Codex and Work mode. Every one of those carries a date you can look up. Older check from 2026-09-11: status.openai.com said fully operational and reported no issues affecting its own systems; history lists Elevated errors affecting ChatGPT Work; 2026-09-03 has reported ChatGPT and Codex multi-component disruption (15 service areas, about 14:56 UTC degraded performance); 2026-09-04 has the Astra paid-user rollout update. Those carry dates too.
Rumour, and what official pages do not say
Calling the 2026-09-03 outage “GPT-6 has kept the whole network down since launch” is not backed by the status page. Calling ChatGPT Work elevated errors “the API is fully stopped” is also false. Unconfirmed: when the next capacity event lands, and whether it stacks with a subscription pause. A screenshot on X shows someone is talking, not that every region is down.
Wait it out, or change roads
Stay on OpenAI until the banner clears
Fits a named incident on the status page when your load can pause for tens of minutes. A poor fit for a 24/7 agent: waiting costs more than swapping models.
Point the same key at another family
Fits lingering HTTP 5xx, or ChatGPT web and the API disagreeing. Point Claude Code at ANTHROPIC_BASE_URL, or Codex CLI at a compatible endpoint, and keep shipping on a model that is not OpenAI.
Three diagnostic steps
Step 1: open status.openai.com and write down the banner plus the latest history title; do not trust only a social screenshot. Step 2: read the HTTP code. 5xx gets exponential backoff; 429 gets lower concurrency; a usage-limit sentence with no HTTP code belongs on the quota page. Step 3: if status is already green and you still fail, check proxy timeouts and key scopes. To change roads, set Claude Code ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN, or point Codex CLI at an OpenAI-compatible base URL.
What to do on QCode, specifically
Sign up and copy the dashboard key. Claude Code: ANTHROPIC_BASE_URL = https://api.qcode.cc/api and ANTHROPIC_AUTH_TOKEN set to that key; walkthrough at anthropic-base-url-setup-guide. Codex CLI: set base_url in config.toml to https://api.qcode.cc/openai. Probe first with gpt-5.6-sol (billed per token, per-model prices are listed on /models), then switch to gpt-6-astra by task; Claude family models go through the Claude Code setup above, since Claude cannot use the OpenAI endpoints and QCode's docs say Claude models only work through the Anthropic Messages API. Next step: one non-streaming tiny request, confirm HTTP 200 and a matching model id in the body.
FAQ
If the ChatGPT website is down, is the API down too?
Not necessarily. The status page splits ChatGPT, Codex, and APIs. Elevated errors affecting ChatGPT Work on 2026-09-10 did not by themselves paint the API component red. Trust the component list on status.openai.com.
How is 503 different from 429?
503 / 502 are usually OpenAI-side capacity or a gateway — back off. 429 is your rate or concurrency — shed load before you change models. Neither means the key was stolen.
Is this the same as Astra burning quota fast?
No. Quota is credits pricing and the five-hour window; see /gpt-6-astra-quota-burn. This page is availability: a red status banner, 5xx, queues, timeouts.
Status is green but my request still fails. What could it be?
Four causes are common. (1) The key only covers some endpoints or lacks the right permissions. (2) A local proxy, CDN, or client timeout is shorter than the model needs to generate. (3) The model had no capacity at that moment; the official error-code guide writes 503 - Model temporarily overloaded. (4) You hit your own rate or concurrency limit; the same guide writes 429 as a reached request rate limit. Even when the overall status reads All Systems Operational, the status page still lists components separately, so one component such as Chat Completions can dip on its own and affect only the matching interface. The official record of 2026-10-14 is exactly such a per-component entry, a degraded-performance note for Codex and ChatGPT Work. Once the status page is green, a global outage is no longer the first hypothesis.
Can QCode make OpenAI recover faster?
No. We cannot add capacity at OpenAI. What we can do is let the same key call a model family that is still answering, so the job finishes.
Codex CLI timed out. Should I upgrade the ChatGPT plan immediately?
Read status and the HTTP code first. A higher ChatGPT tier does not fix 5xx. If it really is quota, change model or billing path — do not start by buying a ChatGPT subscription.
Sources
Live copy: status.openai.com (fetched 2026-10-20, All Systems Operational). Component status: the status page lists every component on its own line, Chat Completions among them. Official error-code guide: the error codes page on developers.openai.com (the official wording for 503 and 429 is quoted in the body). Incident records: the incidents feed of status.openai.com, one entry each for 2026-10-19, 2026-10-15, 2026-10-14 and 2026-10-13. Older entries: status.openai.com (fetched 2026-09-11, fully operational); Elevated errors affecting ChatGPT Work. 2026-09-03 multi-component outage: Crypto Briefing. Astra paid rollout: The New Stack update 2026-09-04. Third-party probe date: llmlatency.dev marks Last verified, dated September 11, 2026.
When OpenAI jitters, keep a path that still runs
One QCode key to move between OpenAI-family and Claude-family models. Plans from ¥63/month; the key is live after payment.
Related
Claude 529 Overloaded
The other vendor's capacity error: exponential backoff, not a prompt rewrite.
Why GPT-6 Astra burns quota so fast
Availability incidents and the credits multiplier are different pages.
Stream closed before completed
Four causes of a cut stream — not the same class as a global 5xx.
Crypto payment guide
Nine coins, 56 networks, zero platform fee for AI coding API credits.
This is a troubleshooting note, not a mirror of OpenAI's official status page. status.openai.com wins; the check date is 2026-10-20. QCode cannot speed up OpenAI's recovery and does not resell ChatGPT subscriptions.