Claude Desktop GPU Process Crash 101457950? Contain It Safely

If Claude Desktop on Windows keeps closing and its log says GPU process gone with exitCode: 101457950, preserve the matching log lines, stop restoring the in-app browser or preview that preceded the crash, update Claude from Anthropic's official channel, and test a plain chat before reopening that browser boundary. The message identifies a crashed Chromium GPU subprocess. It does not, by itself, prove that your physical GPU failed or that one community workaround is safe.

Use this order:

1. Find the exact signature in Claude's own log.

2. Compare it with Windows display-driver and Code Integrity events.

3. Close Claude and contain the in-app browser or preview trigger.

4. Update through Anthropic's supported deployment path.

5. Check the Windows package before using Repair.

6. Validate a plain chat, then cautiously test the suspected trigger once.

Do not start by deleting Claude's package data, disabling Code Integrity, installing an unofficial build, or forcing undocumented Chromium flags. Recent reports show the same exit code on Intel, NVIDIA, and Windows on Arm systems, and several reporters found that driver changes or reinstalling did not prevent recurrence.

Match the Exact Crash

This guide applies when Claude Desktop's main.log contains a line resembling:

GPU process gone
exitCode: 101457950
serviceName: GPU

Some reports also show nearby browser or rendering work such as capturePreviewScreenshot, UnknownVizError, WebGL, or WebGPU messages. Treat those as context, not a universal root cause. Other reports found no reliable trigger.

Use a different diagnosis when:

• PowerShell reports Status : Modified, NeedsRemediation for the Claude package;

• Windows logs display-driver Event ID 4101 at the same time;

• only Claude Code in a terminal fails;

• the app stays open but authorization loops; or

• Windows itself freezes, shows artifacts, or crashes outside Claude.

If the package reports Modified, NeedsRemediation, stop this workflow and use the separate Claude Desktop MSIX repair guide. That page owns package recovery; this page owns the exact GPU child-process crash and containment intent.

Prerequisites

Before testing:

• save or commit irreplaceable project files outside Claude Desktop;

• record the Claude version, Windows version, GPU model, and driver version;

• note the approximate crash time and whether an in-app browser or preview was open;

• close sensitive projects before collecting logs; and

• use an administrator-approved process on a managed device.

Anthropic documents Help > Troubleshooting > Show Logs as the supported way to locate Claude Desktop logs. The PowerShell commands below search the common unpackaged and MSIX log roots without modifying them.

Step 1: Find the 101457950 Signature

Open PowerShell and run:

$claudeLogRoots = @(
  (Join-Path $env:APPDATA "Claude\logs"),
  (Join-Path $env:LOCALAPPDATA "Packages\Claude_pzs8sxrjxfjjc\LocalCache\Roaming\Claude\logs")
) | Where-Object { Test-Path $_ }

Get-ChildItem $claudeLogRoots -Filter "main*.log" |
  Select-String -Pattern "GPU process gone|exitCode: 101457950|capturePreviewScreenshot|UnknownVizError" |
  Select-Object Path, LineNumber, Line

Expected result: PowerShell prints the file, line number, and only matching lines. If $claudeLogRoots is empty, use Claude's Show Logs command rather than guessing another package path.

Copy only the few relevant lines into a private incident note. Review paths, account names, URLs, prompts, tokens, and project details before sharing anything publicly. Do not upload the full log by default.

If the exact exit code is absent, stop using this guide as a diagnosis. A blank window, sign-in failure, package-integrity problem, and Windows display-driver reset need different evidence.

Step 2: Separate Three Failure Boundaries

Check for a Windows display-driver reset during the previous 24 hours:

$since = (Get-Date).AddHours(-24)

Get-WinEvent -FilterHashtable @{
  LogName = "System"
  Id = 4101
  StartTime = $since
} -ErrorAction SilentlyContinue |
  Select-Object TimeCreated, Id, ProviderName, Message

Then check the Code Integrity event mentioned in one recent report:

Get-WinEvent -FilterHashtable @{
  LogName = "Microsoft-Windows-CodeIntegrity/Operational"
  Id = 3033
  StartTime = $since
} -ErrorAction SilentlyContinue |
  Select-Object TimeCreated, Id, ProviderName, Message

Finally, inspect the Claude MSIX package without changing it:

Get-AppxPackage |
  Where-Object PackageFamilyName -Like "Claude_*" |
  Select-Object Name, Version, Status, PackageFullName

Interpret the result narrowly:

One event near another is correlation, not proof of cause. The value of this split is that it prevents a Claude child-process exit from becoming an unnecessary GPU replacement, security-policy change, or destructive package reset.

Step 3: Contain the Browser or Preview Boundary

Several current reports correlate the crash with Claude's in-app browser, browser-pane restoration, or preview screenshot capture. Contain that boundary before changing the machine:

1. Fully quit Claude Desktop.

2. Reopen it without restoring the browser or preview pane if the app offers that choice.

3. Start a new plain chat with no browser, preview, or graphical artifact.

4. Move urgent web research to your normal browser temporarily.

5. Preserve the crashing conversation instead of repeatedly reopening it.

Expected result: a bounded text-only session remains open and no new GPU process gone line appears. This does not fix the underlying defect. It tests whether useful work can continue outside the suspected rendering boundary.

If Claude crashes before a window appears, or it always restores the failing pane, stop repeated launches. Preserve the logs and move to the official update and package checks.

Step 4: Update Through the Official Path

Anthropic distributes Claude Desktop for Windows as an MSIX package. Use the version and update channel approved for your device:

1. Close all Claude processes.

2. Install pending Windows updates that your organization has approved.

3. Update Claude through its supported installer or managed deployment.

4. Restart Windows before testing again.

5. Record the new Claude and driver versions.

Do not assume that updating the GPU driver alone will fix this signature. Current reporters reproduced it after driver changes, GPU selection changes, repair, or reinstall. Updates remain a sensible baseline because they remove known defects, but the observable validation matters more than the act of reinstalling.

If the package status is Modified, NeedsRemediation, use Windows Repair only after every Claude process is closed. Microsoft says Repair retains app data, while Reset permanently deletes app data, preferences, and sign-in state. Keep Reset as a backed-up last resort, not a GPU-crash shortcut.

Step 5: Validate in Increasingly Risky Steps

After restarting:

1. Run the package-status command and confirm Status : Ok when an MSIX package is present.

2. Open Claude Desktop and start one plain text chat.

3. Keep it open for several minutes.

4. Re-run the log search and confirm no new 101457950 entry appeared.

5. Only if you can tolerate another crash, open one innocuous browser or preview page.

6. Re-run both the Claude-log and Windows-event checks.

Expected result: the plain chat remains stable. If the single browser test immediately adds the same line, stop testing that boundary and attach the redacted signature, timestamp, version matrix, and upstream issue link to your support report.

Common Failure Modes

Why Containment Works

Claude Desktop uses Chromium components that run graphics work in a separate GPU subprocess. When that child exits, Claude can report the numeric code even when Windows has not recorded a display-driver reset. Keeping the in-app browser and preview closed removes one graphics-heavy boundary while leaving a plain text workflow available.

The diagnostic sequence works because each step has an observable result:

• Claude's log confirms the exact subprocess signature.

• Windows events distinguish a broader driver or integrity signal.

• AppX status distinguishes a damaged package.

• A text-only session tests containment without destructive changes.

• One bounded browser test checks the suspected trigger after updates.

This mirrors the useful structure of Aident's Ollama network troubleshooting guide: match exact terminology, make the boundary reproducible, state the expected result, and explain why it matters. For agent work that must be isolated from a fragile desktop app, see how to run Claude Code or Codex in a Docker sandbox.

Keep One Read-Only Research Task Moving

While the Claude Desktop browser boundary is contained, you can verify the upstream issue pattern through Aident Loadout without changing Claude's package. Set it up by pasting:

Follow https://aident.ai/SETUP.md

Then ask:

Check my Aident Loadout authentication and Vault status. Find one read-only
GitHub issue-search Action, inspect its current schema, and preflight it.
If the preflight is safe, search for "Claude Desktop GPU process gone
101457950" and return dated issue URLs plus visible reactions and comments.
Do not run a write Action, change a connection, or reveal credentials

Expected result: one schema-checked, read-only search returns current primary issue links outside Claude Desktop. This does not repair the app. It gives you a reproducible support packet while preserving the unstable boundary.

Sources

• Deploy Claude Desktop for Windows, Anthropic Help Center, accessed August 3, 2026

• Claude in Chrome troubleshooting, Anthropic Help Center, accessed August 3, 2026

• Reset or Repair MSIX apps, Microsoft Learn, accessed August 3, 2026

• MSIX troubleshooting guide, Microsoft Learn, accessed August 3, 2026

• Troubleshoot application crashing behavior, Microsoft Learn, accessed August 3, 2026

• Claude Desktop crashes during browser research four to five times daily, opened July 15, 2026

• Browser-pane restore loop reproduces exit code 101457950, opened July 26, 2026

• Code Integrity event 3033 appears near the GPU crash, opened July 26, 2026

• Repeated GPU process crashes interrupt active sessions, opened July 27, 2026

• Three GPU process crashes with no reliable trigger, opened July 28, 2026

Refresh this guide when Anthropic documents a first-party fix, changes the Windows desktop runtime, or closes the active crash reports with a verified resolution.