Claude Code v2.1.149 Adds Per-Category Usage: The New Way to Audit Agent Spending

Usage Is Finally Becoming Attributable

Claude Code v2.1.149 adds a deceptively important feature: /usage now shows a per-category breakdown of what is driving limits usage. The release highlights skills, subagents, plugins, and per-MCP-server cost as separate categories. For teams trying to control AI coding spend, that is more than a quality-of-life improvement.

Most AI coding bills are hard to explain because they arrive as a total. Developers know they used an agent, but they do not know whether the spend came from file reads, browser automation, security review, a background research agent, or repeated test failures. Per-category usage moves the conversation from "Claude was expensive" to "this workflow pattern was expensive."

Why Category-Level Usage Changes Cost Management

Agentic coding is not a single activity. A modern coding session can include a planner, a code editor, a test runner, a documentation skill, a web research skill, and multiple MCP servers. Each component has a different value profile. A security review that finds a critical issue is worth a high token bill. A looping documentation agent that reads the same files ten times is not.

Subagents Are the Biggest Budget Wildcard

Subagents can protect the main context window and speed up research, but they also multiply work. A single prompt can launch several independent investigations, each with its own context and tool calls. If the result is used to make a better decision, that is valuable. If the result duplicates work the main agent was already doing, the team pays twice.

Per-category usage makes it easier to set guardrails. For example, a team might allow subagents freely during architecture planning but discourage them for small edits. Another team might allow web research skills for pricing verification but block them during routine local refactors.

MCP Servers Need Their Own Line Item

MCP servers are powerful because they connect coding agents to issue trackers, docs, databases, browser sessions, and internal systems. They are also a source of hidden context. A broad query to a documentation server can return thousands of tokens. A browser automation server can capture large page states. A repository server can expose more files than the task actually needs.

Once usage is visible per MCP server, teams can ask better questions: which server produces the most tokens, which one produces the most useful answers, and which one should be limited to explicit user approval?

How to Turn /usage Into a Budget Habit

• Check /usage after long coding sessions and record the dominant category.
• Compare subagent-heavy sessions against single-agent sessions for similar tasks.
• Review MCP usage when a task feels more expensive than expected.
• Create team norms for when specialized skills are worth their overhead.
• Estimate high-cost workflows before repeating them across many repositories.

Claude Code v2.1.149 does not lower model prices. It lowers uncertainty. That is often the first step toward lowering actual spend. Use the AI Cost Estimator to translate those usage patterns into model-level cost scenarios.