Claude Desktop on Windows silently spawns a 1.8 gigabyte virtual machine every time you launch the app, consuming roughly 11% of a 16 GB laptop's RAM before you even open a tab. The VM powers Claude's Cowork feature, which lets the AI agent execute code and interact with your desktop. The problem: it starts unconditionally on every launch, not just when you actually use agentic tasks.

The issue gained attention on Hacker News on June 10, 2026, when it accumulated 267 points and 177 comments, revealing that many developers had noticed the Vmmem process consuming memory in Task Manager but had no idea where it came from. A GitHub bug report filed by user tonyrice documented the behavior precisely: a Razer Blade 15 laptop with 16 GB of RAM, no Windows Sandbox enabled, and only VirtualMachinePlatform virtualization active. Yet on every launch, the Hyper-V Host Compute Service (vmcompute) fires up and spawns a full virtual machine.

When Claude Desktop starts on Windows, it triggers the Hyper-V Host Compute Service via an RPC interface event. That service spins up a vmwp.exe process that hosts a full virtual machine, which appears in Task Manager as Vmmem. The memory consumption is consistently around 1.8 GB, regardless of whether you ever open the Cowork tab or intend to run any agentic task.

Why Does Claude Need a Virtual Machine at All?

The design rationale is sound. Local agent sandboxing requires isolation. When Claude can read your filesystem, execute shell commands, and interact with running applications, you need a hard boundary between the agent and the host. A virtual machine provides that boundary more reliably than process-level sandboxing. Anthropic is not wrong to use one.

The problem is the always-on startup behavior. One Hacker News commenter captured the frustration: "The VM itself is for Claude Cowork which does all work within the VM sandbox. That doesn't help answer why they spin it up immediately and don't have a way to disable it though." Several users on the GitHub issue confirmed they had the same Vmmem behavior even with a work account where Cowork was explicitly disabled at the organizational level.

Who Gets Hit Hardest by the Memory Drain?

The impact varies by user profile, but certain groups feel the cost most acutely. The memory overhead is particularly sharp on resource-constrained machines, while corporate environments face additional friction from IT policy restrictions.

• 16 GB RAM Laptop, Chat-Only Use: The VM consumes roughly 11% of total RAM before you open a single tab, leaving less headroom for your actual work.
• Corporate Machine with Hyper-V Policy Disabled: VM launch fails entirely, producing repeated errors in the Hyper-V Compute Admin event log with the message "The specified property query is invalid: The virtual machine or container JSON document is invalid."
• Machine with Full System Drive: One Hacker News commenter noted the VM comes with a roughly 10 GB VM bundle on disk that cannot be easily removed, in addition to the runtime RAM cost.
• Developer Running Docker and WSL2 Simultaneously: Memory pressure from multiple competing hypervisors creates resource contention that degrades performance across all tools.
• User with VirtualMachinePlatform Disabled Entirely: The VM does not start, and the Cowork feature becomes unavailable.
• macOS User: A related issue (number 30972) reports that the Apple VirtualMachine process loads at startup at roughly 2.61 GB.

The corporate IT angle is particularly sharp. Many enterprises block or disable Hyper-V features via Group Policy for security and licensing reasons. On those machines, the VM launch fails and produces repeated errors in the Hyper-V Compute Admin event log. One user also reported finding 2,689 stale session files in the local-agent-mode-sessions folder, named in Docker-style format (nifty-dreamy-volta, tender-vigilant-goodall), accumulated from previous Cowork sessions that were never cleaned up.

How to Manage the VM Overhead and Cowork Feature

• Monitor Task Manager Regularly: Open Task Manager and search for the Vmmem process to confirm the VM is running and consuming memory. Document baseline memory usage before and after Claude Desktop launches to quantify the impact on your specific machine.
• Check Your Cowork Permissions: If you work in a corporate environment, verify whether Cowork is explicitly disabled at the organizational level. Even if disabled, the VM may still launch, so confirm with your IT department whether Hyper-V policies are blocking the feature.
• Clean Up Stale Session Files: Navigate to %APPDATA%\Claude\local-agent-mode-sessions\ and delete accumulated session files. Note that deleting these files and manually killing VM processes will not prevent the VM from respawning on the next app launch.
• Evaluate Alternative Workflows: If you primarily use Claude for chat and do not rely on agentic features, consider whether the memory cost justifies the feature availability. Claude Code, the terminal-first alternative, does not bundle the VM at all.
• Report Issues to Anthropic: If you experience unexpected VM behavior or errors, file a detailed bug report on the anthropics/claude-code GitHub repository with your system configuration, RAM, disk space, and virtualization settings enabled.

What's Changing With Claude's Naming and Model Tiers?

Beyond the VM issue, Anthropic is also reshaping how developers think about Claude's model lineup. The company introduced a new naming structure with Claude Fable 5, breaking from its previous tier hierarchy. For most of Claude's public history, the naming system was straightforward: Haiku for fast, cheap tasks; Sonnet for middle-ground reasoning; and Opus for flagship capability. Version numbers tracked iteration within each tier.

With Claude Fable 5, Anthropic introduced something structurally new: a named capability class called Mythos, positioned above Opus. Fable is the first model in that class. The tier name is now a class name, not a poetic register. Haiku, Sonnet, and Opus are all forms of writing that describe something about the model's style. Mythos is a category that describes what the tier is for, not what it sounds like.

The top tier now has a model name distinct from the class name. Fable 5 lives inside Mythos class the way a product lives inside a product line. The model ID claude-fable-5 does not contain the word "Mythos." Mythos is the capability bucket; Fable is the named flagship inside it. This separation lets Anthropic release a second Mythos model, say one optimized for agentic tasks or with different safety characteristics, without it being "Fable 5.1".

Mythos is not just a marketing label. Based on the Fable 5 release details, the class comes with distinct operational characteristics. Pricing is in a different band: at $10 per million input tokens and $50 per million output tokens, Fable 5 is roughly 2x the cost of Opus 4.8 on both input and output. The context window extends to 1 million tokens, which puts it in a different class of problem. Tasks that require holding an entire codebase, a full legal document set, or a long conversation history in context are now viable in a single call rather than requiring chunking logic.

Safety and evaluation requirements at the Mythos tier are understood to be more rigorous than at Opus. Anthropic's internal AI Safety Level (ASL) framework gates model releases on evaluation benchmarks that grow stricter as capability increases. A model above Opus is, by that framework, a model that required new gates to clear. What those gates look like for external API users in terms of rate limits, use-case restrictions, or enterprise agreements is something worth watching as the tier matures.

The naming shift sparked humor in the developer community. A satirical table called "Anthropic's Model Naming, Extrapolated" hit 204 points on Hacker News, joking that the convention now stretches far enough to extrapolate a full literary canon: Saga, Canon, Lore, Cinematic Universe, all the way to "Zack Snyder's Saga (terminal turns black and white, becomes harder to follow)." The joke landed because the underlying observation is real: with Claude Fable 5, Anthropic crossed a line it had not crossed before.