Screenshot from GitHub Releases (github.com/anthropics/claude-code) Claude Code 2.1.174: VS Code Now Shows Exactly Where Your Tokens Went
Today's Claude Code update adds a detailed usage attribution breakdown to the VS Code /usage dialog, showing cache misses, long context sessions, subagent use, and per-skill, per-MCP token consumption over the last 24 hours or 7 days.
Anthropic pushed Claude Code 2.1.174 early on June 12. Most of the release is bug fixes, but one new feature stands out for anyone trying to understand what’s actually consuming their token budget: a detailed usage attribution breakdown in VS Code.
Token Attribution in the /usage Dialog
Opening /usage in VS Code now shows a breakdown of what drove your consumption over the last 24 hours or the last 7 days. The categories are:
- Cache misses — requests that couldn’t be served from the prompt cache
- Long context — sessions that exceeded the standard context window
- Subagents — tokens consumed by spawned sub-agents
- Per-skill, per-agent, per-plugin, per-MCP — line items for each connected component
This answers a question that comes up constantly for heavy users: “Where is all my context going?” Before this, you’d see total consumption but no breakdown by source. If you had a lot of subagents running, or one MCP server being called repeatedly, or an expensive skill that was triggering constantly, you had to guess. Now you can see it.
The 24h/7d toggle is useful for spotting patterns. A spike in one day versus a week of data tells a different story.
Model Picker Fix
The /model picker had a display bug where it hid the model family that Default resolves to. On Max, Team Premium, and Enterprise plans, Default resolves to Opus, but Opus wasn’t showing as its own row. On Pro and Team plans, where Default resolves to Sonnet, the same problem existed. That’s fixed now — whatever Default maps to appears as its own selectable row.
A related fix: if you’d pinned a specific Sonnet version via ANTHROPIC_DEFAULT_SONNET_MODEL, the model picker would still show a hardcoded Sonnet label rather than the pinned version. That’s corrected too.
Enterprise Billing Banner Fix
Enterprise accounts with usage-based billing were seeing an incorrect “Fable 5 is now consuming usage credits” notification. The banner was triggered by a logic error that didn’t account for enterprise billing mode. It’s been removed for accounts where it shouldn’t appear.
Bedrock GovCloud Fix
AWS GovCloud regions (us-gov-east-1, us-gov-west-1) were getting the wrong inference profile prefix in Claude Code’s Bedrock integration. The code was using global instead of us-gov, which caused 400 errors when trying to derive model IDs. Fixed in this release.
Skill Hot-Reload Optimization
When a single skill file changed, Claude Code was re-sending the entire skill listing to reannounce it. That’s been tightened up so only the changed skill gets re-announced. For projects with many skills, this reduces unnecessary overhead every time you edit a skill file.
v2.1.175
A second release shipped the same day, adding the enforceAvailableModels managed setting for enterprise environments. When this is enabled, the availableModels allowlist also constrains what Default can resolve to: if Default would pick a model not on the allowlist, it falls back to the first allowed model instead. It also prevents user or project-level settings from widening an allowlist set by a manager.
That’s aimed at organizations that need to enforce consistent model usage across their teams and don’t want individual users overriding the approved list.