ChatGPT Web crashes/freezes when loading specific long chats (Client-side)

Description of the Issue

When attempting to load and view a specific chat conversation on the ChatGPT web application, the browser tab becomes unresponsive, lags severely, and frequently crashes completely (e.g., browser out-of-memory / page crash).

Even in regular or moderately sized chats, there is noticeable stuttering during scrolling and typing, pointing toward a broader client-side DOM rendering or memory leak issue.


Environment & Reproduction Details

  • Affected URL / Chat ID format: Specific conversation threads (e.g., long markdown/code/artifact history).
  • Devices Tested: Multiple distinct machines (Desktop & Laptop with high-end hardware configurations).
  • Browsers Tested: Chrome / Edge / Firefox (Hardware acceleration enabled).
  • Scope: Happens across multiple devices independently, ruling out local machine hardware limits or local cache corruption.

Symptoms & Observed Behavior

  1. High Memory / CPU Spike: Upon opening the conversation thread, tab memory usage spikes rapidly until the browser tab crashes.
  2. UI Freezes & Stutter: Severe stuttering when scrolling through previous turns or attempting to focus the text input box.
  3. Rendering Bottleneck: Client-side rendering seems to struggle virtualizing or mounting long message histories or complex markdown blocks.

Expected Behavior

Long conversation histories should be virtualized, lazily loaded, or paginated so that client-side memory usage remains stable and does not crash the browser tab.


Has anyone else encountered severe tab crashes on specific long chats recently, or is there a known workaround to truncate/export the thread without triggering a tab crash?

I can add measured Firefox Profiler evidence for the same symptom family: very high CPU in a long conversation, delayed or blank sections while scrolling upward, multi-second input and paint delays, and extremely slow visible message updates.

Environment

  • ChatGPT Pro on ChatGPT Web
  • Firefox 153.0 on Linux Mint, X11/GTK
  • AMD Ryzen 9 9950X, 16 physical / 32 logical CPUs
  • 22 August 2026, Europe/Zurich (CEST)
  • A very long conversation with substantial Markdown and structured content

Reproduction

  1. Reload the affected long conversation.
  2. Scroll upward while older sections are progressively restored.
  3. Observe sustained CPU use, delayed input/paint, and areas that remain blank for seconds before appearing.
  4. Continue the conversation and observe very slow visible response updates.

The two strongest retained intervals measured as follows:

Scenario Retained span Main-thread CPU One-core utilization Explicit long tasks Minor-GC markers Style markers
Reload and upward scroll 45.361 s 36.738 s 80.99% 6 totaling 1.560 s; max 648 ms 184,032 2,493
Continue and response rendering 67.199 s 48.606 s 72.33% 310 totaling 45.358 s; max 246 ms; p95 196 ms 120,836 7,625

In the reload/scroll capture, first-party ChatGPT JavaScript appeared somewhere in the stack for 93.90% of sampled main-thread CPU, while extension JavaScript accounted for 0.22%. The continuation capture also recorded 7,102 container-query style updates, 3,096 paint/display transactions, and 678 synchronous-reflow markers.

Sampled inclusive stacks repeatedly passed through ChatGPT conversation/display-state work, synchronous state updates, and display-item derivation. The stacks overlap, and minified bundles do not identify one component as the sole root cause. The evidence does show enough first-party main-thread work to delay event handling and painting even on a high-core-count desktop.

Requested investigation / expected behavior

  • Window or unmount off-screen turns while preserving scroll position.
  • On reopen, show completed user-visible reasoning summaries, tool logs, and activity details as lightweight collapsed placeholders; fetch/render them on click and release their DOM/AST/highlighter state when collapsed again.
  • Store a safe content-hashed, renderer-versioned presentation representation for completed messages server-side so a page load does not require every client to reparse all historical raw Markdown.
  • Maintain Markdown parser checkpoints and open-block state. For appended content, reparse from the nearest affected opener or safe block boundary rather than automatically scanning from the beginning of the message.
  • Buffer incoming deltas and commit them at a bounded adaptive cadence, such as 100–500 ms, with an optional 500–1000 ms low-CPU mode for slower systems. Stop and input controls must remain immediate.
  • Freeze completed blocks and avoid recomputing a full display list for a small tail update.
  • Audit synchronous state flushes, style/container-query churn, observers, content-visibility transitions, and scroll anchoring.
  • Keep scrolling, Stop, and navigation responsive while older turns load.

I am not claiming that React alone is the bug, that every completed message is reparsed for every token, or that switching transport to WebSockets would fix rendering. The measured defect is first-party main-thread scaling under long and streaming conversations.

Hi @subhaan0804 and @Orry ! Thanks for sharing all these details, especially the profiling data and troubleshooting you've already done.

Just wanted to check if you're still running into freezing, lag, or browser crashes when opening longer conversations.

If it's still happening, could you let me know if you've noticed any improvement recently or if it's still affecting the same conversations?

You can also check out the ChatGPT performance troubleshooting guide for additional guidance.

-Basil

I’m also affected, with Mozilla based browsers, since several weeks

Hi @ik314 , thanks for confirming you're seeing this too.

Since this has been happening for a few weeks, could you reach out to OpenAI Support so they can take a closer look?

When you contact Support, please include your browser and version, OS version, an approximate timestamp, steps to reproduce the issue, and a screen recording if possible.

If you can reproduce it, it would also be helpful to include a HAR file so Support can review the browser's network activity and request timings. You can find the instructions under Recording a HAR file for Troubleshooting UI Issues.

Once you open the Support request, feel free to share the case number here and I can take a look at the details provided privately.

-Basil