Claude Code's Parallel Work Mode Lets AI Juggle Tasks—But Tokens Disappear Fast

Anthropic is shipping a revamped Projects system that turns Claude Code into a workforce coordinator, spawning multiple parallel sessions to tackle subdivided engineering goals—though the speed gains come with a catch for users watching their token budgets.

How the New Coordinator Works#

The enhanced Projects feature, now in beta for select Claude Pro and Max subscribers, introduces a supervisor layer that accepts high-level development objectives and automatically decomposes them into independent subtasks. Rather than forcing developers to manually spin up separate Claude Code sessions and stitch results together, a single "coordinator" Claude instance delegates work across multiple parallel threads, each running in its own isolated cloud environment with a dedicated repository branch.

Think of a team retiring a legacy API endpoint across multiple client libraries: the coordinator could launch parallel threads to update callers in the web client, mobile app, and backend simultaneously—each running tests and opening pull requests before identifying merge dependencies. Developers monitor progress through the main Project interface or zoom into individual threads to intervene without manually orchestrating every session.

Each worker thread can further subdivide its assignment using Claude Code's existing subagents, loops, and workflows, creating nested layers of delegation when needed.

The Token Cost Problem#

This efficiency comes with a significant trade-off: each parallel thread counts as a full Claude Code session against a subscriber's usage plan. Running multiple threads simultaneously accelerates token consumption and can exhaust monthly allocations faster than traditional serial workflows. Anthropic openly acknowledges this dynamic, noting that Projects will "hit usage limits sooner when multiple threads are running."

The timing raises questions as Anthropic defends a class-action lawsuit filed by Max plan subscribers who allege the company advertised usage increases without clearly disclosing the weekly rate-limiting caps that kicked in.

Shared Context Across Threads#

Anthropic is layering in a shared memory system so knowledge gained during one thread's execution propagates to others automatically. This reduces redundant context-passing: if one thread discovers a changed release date or learns which team member approves certain services, that insight carries forward as subsequent threads tackle their portions.