ChatGPT is usually not working because of one of four failures: OpenAI's servers are overloaded or down, your login session has expired, the network path between your device and OpenAI is blocked or degraded, or the browser/app state is corrupted. Outages, expired tokens, and stale caches account for most incidents. Distinguishing a service-wide outage from a local problem takes about 30 seconds, and the fix differs sharply for each.

What an outage actually is

When ChatGPT fails, the failure usually originates far from your keyboard. OpenAI's chat product is not one server but a chain: edge/CDN layer, authentication service, API gateway, model inference cluster, and a moderation/streaming layer that pushes tokens back to your browser. A break anywhere in that chain produces the same symptom from your side: a spinner, a "network error," or a half-written reply that stops mid-sentence.

The most common real-world cause is inference capacity, not a crashed data center. Language models run on GPU clusters with finite throughput. When request volume exceeds provisioned capacity, the queue backing up shows up to users as slow first-token latency or a 429 "too many requests" error. OpenAI has publicly acknowledged capacity constraints during high-demand periods, and its status page distinguishes degraded performance from full outages.

A second frequent mechanism is authentication expiry. ChatGPT sessions rely on short-lived access tokens refreshed by longer-lived refresh tokens. If the refresh fails—because of clock skew, a blocked cookie, or a changed password that invalidates old sessions—the UI still loads, but every message returns an authorization error. This looks like "ChatGPT is broken" when the model service is perfectly healthy.

A third mechanism is streaming interruption. Replies are delivered over a persistent connection. Corporate proxies, VPNs, and some mobile carriers terminate long-lived connections aggressively, so the model generates an answer but your client never receives the tail. The result is a reply that hangs or truncates at a consistent point—a signature of a network middlebox, not a model failure.

Plainly: most "ChatGPT is down" reports are capacity queues, expired sessions, or connections killed in transit, not a dead server.

How often failures happen, and what the numbers mean

Published uptime figures vary by measurement method, so treat any single number with caution. Status-page uptime and independent third-party monitors disagree because they sample different endpoints at different intervals. What matters is the order of magnitude and the failure profile.

Failure type Typical user symptom Detection speed Who can fix it
Platform outage Error page, spinner on all devices Seconds to minutes OpenAI only
Capacity/throttling 429 errors, slow first token Minutes OpenAI (wait) or plan tier
Auth/session expiry Login loop, "unauthorized" Immediate You (re-login)
Network/proxy block Hangs, truncation, timeouts Immediate You or your IT team
Local cache/cookies Broken UI, stale errors Immediate You (clear cache)
Account/billing state Feature locks, model limits Immediate You (check plan)
Regional restriction Hard block in some countries Persistent Neither (policy)

The practical takeaway: a genuine platform-wide outage affects everyone and resolves without your intervention. If your issue persists while other users report normal service, the cause is almost certainly local.

A counterintuitive detail experts rely on: error codes are more diagnostic than error messages. A 401 is authentication, a 403 is permission or policy, a 429 is rate limiting or capacity, a 500-series is server-side, and a generic "network error" with no code is usually your connection. Users read "something went wrong"; the underlying status code tells you where to look.

Edge cases that mimic outages

Some failures look identical to an outage but have distinct causes.

Red-flag warning: If ChatGPT breaks immediately after you change your password or enable new security settings, suspect session invalidation, not an outage. Red-flag warning: If replies consistently truncate at the same length on one network but complete on cellular, a proxy or VPN is cutting the stream. Red-flag warning: If a paid feature (higher limits, certain models) suddenly vanishes, check billing and plan status before assuming downtime. Red-flag warning: If the app fails but the browser works, the problem is local app storage, not the service. Red-flag warning: If login loops endlessly, third-party cookie blocking or an ad blocker is often the culprit.

Regional restriction is its own category. ChatGPT has limited availability in some countries, and a hard block there is policy, not malfunction—no cache clear will fix it. Similarly, enterprise SSO configurations can cause login failures that look like server errors but are identity-provider misconfigurations.

Mobile carriers add another layer. Some networks use carrier-grade NAT and aggressive connection timeouts that break long polling or streaming, so mobile sessions fail where Wi-Fi succeeds. This is a network path problem, not an OpenAI problem.

Misconceptions worth discarding

The single most widespread misconception is that "ChatGPT is down" always means the model is down. Usually the model is fine; the plumbing between you and it is not.

A second misconception is that clearing your cache is a universal fix. It helps with local UI and session corruption, but it does nothing for a capacity queue, a proxy block, or a regional restriction. Applying it indiscriminately wastes time and obscures the real cause.

A third is that a status page showing "operational" proves your issue is local. Status pages lag and may not capture partial degradation on specific model endpoints or regions. Absence of a reported incident is weak evidence, not proof.

A fourth is blaming the model itself when replies feel "dumber." Quality shifts usually reflect model version changes, context-window limits, or system instructions, not server failure. Treating a reasoning change as an outage leads people to restart endlessly when nothing is broken.

Plainly: the error type, not the error text, tells you whether to wait, re-login, or call IT.

Practical actions, with thresholds

Work from the widest scope to the narrowest. Each step resolves a different failure class.

Check service status first. Visit OpenAI's status page and an independent monitor. If both show a confirmed incident, wait—most incidents resolve within minutes to a few hours. Do not troubleshoot locally during a confirmed platform outage. Test across scopes. Try the same prompt on a second device and a different network (Wi-Fi vs. cellular). If it fails everywhere, suspect the service or your account; if it fails only on one setup, suspect local network or app state. Fix authentication. Log out completely, clear cookies for the site, and log back in. If you use SSO, confirm your identity provider is healthy. A 401 or login loop points here. Clear local state. Remove app cache or reinstall the mobile app if the browser works but the app does not. This targets corrupted storage, not the model. Bypass the middlebox. Disable VPN, corporate proxy, or ad blockers and retest. If truncation stops, a network intermediary was cutting the stream. Ask IT to allow persistent connections to OpenAI endpoints. Address throttling. A 429 means you hit rate limits or capacity is saturated. Wait, reduce request frequency, or check whether your plan tier supports the model you're using. When in doubt, one retry after 30-60 seconds distinguishes a transient burst from a persistent limit.

Reconcile the thresholds explicitly: a confirmed status incident means wait; a 401/403 means fix authentication; a 429 means rate/capacity; a persistent failure only on one network means fix the path. Do not apply the cache-clear step until you've ruled out a confirmed outage and a path problem, or you will keep treating symptoms.

Limits of what can be concluded

You cannot diagnose a platform problem from a single device, and you cannot diagnose a local problem while others report a confirmed outage. Both directions produce false certainty. Uptime statistics are inconsistent across sources, incident reports lag reality, and many failures are partial and endpoint-specific, so absence of evidence is not evidence of health.

If a problem persists after you've confirmed the service is up, tested a second network, re-authenticated, and cleared local state, the remaining possibilities—account state, billing, enterprise policy, or a region block—require information only OpenAI support or your IT administrator has. No amount of local troubleshooting can resolve a policy restriction or an account flag.

The honest boundary: mechanisms are well understood, but specific root causes of any single incident are known only to the operator, and dated measurements should not be read as current.