ChatGPT Keeps Crashing: Every Cause and Fix Worth Trying

If ChatGPT keeps crashing on you, you're dealing with one of roughly three things: a server-side outage you can't fix, a client-side conflict you can, or a network-layer problem sitting quietly between the two. This guide covers all three categories with specific steps, ordered by likelihood. No fluff, no "have you tried turning it off and on again" padding. Just the diagnostics.

Key Takeaways

• Check a third-party status tracker first. The official status page sometimes never acknowledges real incidents that users are actively experiencing.
• Most "freezing" isn't a crash. It's a broken streaming connection, often caused by ad blockers, VPN interference, or WebSocket domains being silently blocked by your network.
• A hard refresh (Shift + Reload) and clearing site data (not just cache) resolves a surprising share of stuck-loading and blank-page issues.
• The Windows desktop app has a specific failure mode: the process stays alive in Task Manager after a silent crash, blocking any clean relaunch until you force-quit it.
• If your system clock is off by more than a few minutes, TLS certificate validation can fail entirely, producing login loops or connection errors that look like crashes.

Is It Actually Down, or Is It Just You?

It might genuinely be down. On April 20, 2026, a global outage lasted roughly 90 minutes, with over 13,000 concurrent reports at peak. That's not a browser cache problem. And that wasn't a one-off: independent monitoring logged 59 separate incidents in a recent 90-day window, one major and 58 minor, with a median duration of about an hour and 47 minutes.

Your first step should always be checking a third-party status tracker. The official status page is useful but incomplete. Some real service disruptions confirmed by multiple user reports are never acknowledged on the official page at all. Sites like Pulsetic, UptimeRobot, and Entireweb aggregate user reports independently.

If the trackers show a spike in reports matching your timeframe, close the tab and wait. Nothing you do locally will help during a server-side outage. A confirmed worldwide outage in early 2025 resolved in under 50 minutes once the fix was deployed. Most incidents follow a similar arc. Go get coffee.

Why Does ChatGPT Freeze Mid-Response Instead of Showing an Error?

Because it streams responses word-by-word over a persistent connection. This is architecturally different from a normal web request where you ask, wait, and get a complete response back. If the internet connection drops or introduces packet loss during that stream, the interface just stops. No error message. No timeout dialog. It looks frozen, but it's actually a severed connection that the UI doesn't gracefully handle.

This is why "ChatGPT froze" and "ChatGPT crashed" feel identical from the user's perspective but have completely different root causes. A crash implies the application broke. A freeze usually means the network pipe broke. The fix depends on which one actually happened.

What should you do when a response hangs?

Wait 30 to 60 seconds, then click "Stop generating" followed by "Regenerate." This resets the streaming connection without losing your conversation context. If the response was partially rendered, you'll usually get a clean completion on the retry. If it hangs again on the same prompt, the problem is almost certainly on your side or in the network layer, not on the server.

How Do You Fix the Blank Page or Stuck Loading Screen?

A standard page reload often doesn't cut it here. The official troubleshooting guidance recommends a hard refresh (Shift + Reload), clearing site data specifically for chat.openai.com (not just the browser cache globally), or opening an incognito window with all extensions disabled.

The "clearing site data" part matters more than people realize. Browsers cache service worker registrations, IndexedDB entries, and stale authentication tokens that survive a normal cache clear. In Chrome: open DevTools, go to the Application tab, click "Clear site data" with all checkboxes selected. In Firefox: go to Settings, Privacy, Manage Data, search for the domain, and remove everything.

After clearing, sign out and back in. This forces a fresh authentication handshake and new session tokens. If the problem was a corrupted or expired token sitting in local storage, this kills it.

Are Your Privacy Tools Causing the Crashes?

Probably, if you're running any. This is the frustrating intersection of privacy-preserving software and cloud AI's real-time streaming architecture. They aren't built to cooperate.

Ad blockers, privacy-oriented browser extensions, and network-level DNS filtering tools are repeatedly identified as causes of freezing and connection failures. The mechanism varies by tool:

• Ad blockers can intercept or block the WebSocket connections used for response streaming. The page loads fine, but the moment a response starts streaming, the blocker kills the connection.
• Privacy extensions like uBlock Origin or Privacy Badger may block third-party domains that the application depends on for authentication or real-time communication.
• Some users have traced freezing specifically to a WebSocket domain being blocked by network security tools. Unblocking that domain resolved the issue completely.

The diagnostic is simple: open an incognito or private browsing window (which disables extensions by default in most browsers) and try loading a conversation. If it works cleanly in incognito, one of your extensions is the culprit. Disable them one by one until you find it.

If you're using a system-wide ad blocker or DNS filter (Pi-hole, NextDNS, AdGuard Home), check your query logs for blocked domains during a freeze. You'll usually find the smoking gun there.

Can VPNs and Corporate Proxies Cause ChatGPT to Crash?

Yes, and the failure mode is subtle. VPN and proxy interference can intercept the authentication handshake, causing login loops. You click sign in, the page redirects, and you end up back at the login screen. Repeat indefinitely. This isn't a crash in the traditional sense, but it's indistinguishable from one if you don't know what you're looking at.

Corporate proxies with SSL inspection are particularly bad here. They terminate your TLS connection, inspect the traffic, then re-encrypt it with their own certificate. If the proxy's certificate authority isn't trusted by your browser or if the proxy doesn't handle WebSocket upgrades correctly, the streaming connection either fails silently or gets terminated mid-response.

Try disconnecting from the VPN or switching to a different server. If you're on a corporate network with a mandatory proxy, talk to your IT team about whitelisting the relevant domains.

What about system clock drift?

An out-of-sync system clock can invalidate the TLS certificate check needed to connect securely. TLS certificates have a validity window. If your system thinks it's 2024 or 2028, the certificate looks expired or not-yet-valid, and the connection fails. This is rare on modern systems with automatic time sync, but I've seen it happen on dual-boot Linux machines where the hardware clock gets confused, or on VMs where the guest clock drifts from the host.

On Windows: Settings > Time & Language > Date & Time > toggle "Set time automatically" off and back on. On macOS: System Settings > General > Date & Time > enable "Set date and time automatically." On Linux: sudo timedatectl set-ntp true.

Why Does the Windows Desktop App Keep Crashing?

The Windows desktop app has a specific, well-documented failure pattern. The app crashes silently, but its process keeps running in Task Manager, blocking a clean relaunch. You click the icon, nothing happens. You click it again, nothing happens. The app isn't starting because the previous instance's process is still alive in the background, holding its lock.

The fix:

1. Open Task Manager (Ctrl + Shift + Esc).
2. Find the process in the list. It may appear under "Background processes" rather than "Apps."
3. Right-click it and select "End task."
4. Relaunch the app.

If this happens repeatedly, the app installation might be corrupted. Uninstall it from Settings > Apps, restart your machine, and reinstall from the Microsoft Store.

Around June 26 to 27, 2026, users reported a broken UI update where sidebar icons vanished, conversations appeared empty, and the login button stopped responding, all tied to a Microsoft Store push update that shipped before the client cache was invalidated. If you're hit by something like this, clearing the app's local data or waiting for a subsequent patch is usually the only path. These kinds of issues resolve within a day or two as a hotfix rolls out.

Does Browser Choice Matter?

More than you'd expect. Some crash and freeze reports trace back to browser-specific causes, including a CSS or pointer-events bug that locks the input box entirely. You can see the interface, scroll through old conversations, but the text input field won't accept focus. Clicking it does nothing. This isn't a network issue or a server issue. It's a rendering bug.

If you're experiencing input-box lockups or visual glitches:

• Try a different browser entirely. If it works in Firefox but not Chrome (or vice versa), you've isolated it.
• Disable hardware acceleration in your current browser. In Chrome: Settings > System > "Use hardware acceleration when available" > toggle off, then restart the browser. GPU driver bugs can cause rendering issues that look like application crashes.
• Check if your browser is up to date. Outdated browsers can have stale Web API implementations that break modern streaming features.

What If None of the Local Fixes Work?

If you've tried a different browser, a different network, disabled all extensions, cleared all site data, confirmed your clock is synced, and the problem persists, you're probably dealing with a server-side issue that isn't showing up on the status trackers yet. Or your account specifically has a problem.

The official guidance at this point is to collect a HAR file and browser console logs before contacting support. A HAR file captures the full sequence of network requests, including failed ones, with timing data. It's the single most useful artifact for diagnosing what's actually breaking.

To capture a HAR file in Chrome:

1. Open DevTools (F12 or Ctrl + Shift + I).
2. Go to the Network tab.
3. Check "Preserve log" so requests aren't cleared on navigation.
4. Reproduce the crash or freeze.
5. Right-click anywhere in the network request list and select "Save all as HAR with content."

Also open the Console tab and screenshot or copy any red error messages. Include the conversation URL and your browser version in the support ticket. This saves two rounds of back-and-forth with support asking for information you could have provided upfront.

One note on HAR files: they contain your session cookies, authentication tokens, and the full content of every request and response captured during the recording. If privacy matters to you (and you're reading a blog post on a privacy-focused AI assistant's site, so presumably it does), review the HAR file before sending it. Strip or redact anything you're not comfortable sharing.

Why Do Feature Launches Cause Crashes?

When major updates launch, traffic surges and the system gets overloaded, producing temporary rate limits, slow loading, or widespread errors. This is a predictable pattern at this point. Big product announcements reliably correlate with degraded service for the next few hours.

There isn't a fix for this. It's a capacity problem. If you notice a crash coinciding with a product announcement or a trending topic about a new feature, that's probably why. Wait an hour or two for the traffic to normalize.

What Does This Pattern of Outages Actually Mean?

59 incidents in 90 days is roughly one incident every 36 hours. Most are minor. Most resolve quickly. But if you're relying on a single cloud AI service for work that matters, "usually works" and "always works" are meaningfully different.

This isn't an argument against cloud AI. It's an observation about concentration risk. When one service handles your conversations, your memory, your context, and your workflow, a 90-minute outage isn't just an inconvenience. It's a full stop. Your context is gone. Your history is inaccessible. Your work pauses until someone else's infrastructure comes back.

We built Selina around the assumption that this kind of dependency is worth thinking about. Selina remembers you across conversations, with memory encrypted at rest and routed through a stack of frontier models per task. It doesn't solve the fundamental physics of cloud infrastructure (anything hosted remotely can go down), but distributing across multiple providers at the routing layer means a single provider outage doesn't necessarily take you offline.

Quick Reference: Crash Symptoms and First Steps

• Blank page or infinite loading spinner: Hard refresh (Shift + Reload), then clear site data for the domain specifically, then try incognito.
• Response freezes mid-generation: Wait 60 seconds, click "Stop generating," then "Regenerate." If persistent, check for VPN or ad blocker interference.
• Login loop (redirects back to sign-in): Disable VPN, check system clock, clear all cookies for the domain.
• Input box won't accept text: Try a different browser. Disable hardware acceleration. Check for CSS-related rendering bugs.
• Windows app won't launch: Open Task Manager, end the background process, then relaunch.
• Everything works in incognito but not in your normal browser: An extension is the cause. Disable them one by one to isolate it.
• Nothing works, including different browsers and networks: Check third-party status trackers. If it's not showing as down, collect a HAR file and contact support.

Is There a Way to Prevent Future Crashes?

You can't prevent server-side outages. You can reduce client-side friction significantly by doing a few things once and not thinking about them again:

1. Whitelist the relevant domains in your ad blocker and network-level DNS filter. Add exceptions rather than disabling the tools entirely.
2. Keep your browser updated. Stale browser versions are a consistent source of WebSocket and streaming compatibility issues.
3. If you use the Windows desktop app, pin Task Manager to your taskbar. You'll need it eventually.
4. Bookmark a third-party status tracker. When something breaks, check there before spending 20 minutes debugging your own setup.
5. If your work genuinely depends on AI availability, don't depend on a single service. Have a fallback.

If you want to try a fallback that keeps your conversation memory intact across sessions, start a free 7-day trial of Selina, no card required.

Michael C.

Founder & Principal Engineer, Selina Labs

Michael builds Selina, a privacy-first AI that remembers you across conversations. He ships security-sensitive AI in production — real attacks, real fixes, measured in minutes and dollars — and writes about privacy, security, and LLMs from that seat. Top Rated Plus and expert-verified on Upwork.

• www.upwork.com
• x.com
• medium.com
• selinalabs.io