Bug: Single trivial message consumes ~17.4k tokens

Resolved 💬 3 comments Opened Apr 8, 2026 by amryadam Closed Jun 14, 2026

Bug: Single trivial message consumes ~17.4k tokens

Model: Opus 4.6 (1M context) — claude-opus-4-6[1m]

The Message

User: today is ...
Assistant: Today is April 8, 2026 (Wednesday).

That's it. Two short lines. Yet it consumed ~17.4k tokens.

Before (session start)

Total Usage: 17.8k / 1M tokens (2%)

| Category | Tokens | % of Context |
|----------|--------|-------------|
| System prompt | 6.1k | 0.6% |
| System tools | 6.1k | 0.6% |
| Custom agents | 247 | 0.0% |
| Memory files | 385 | 0.0% |
| Skills | 3.4k | 0.3% |
| Messages | 1.6k | 0.2% |
| Free space | 961.2k | 96.1% |
| Autocompact buffer | 21k | 2.1% |

After (one message exchanged)

Total Usage: 35k / 1M tokens (3%)

| Category | Tokens | % of Context |
|----------|--------|-------------|
| System prompt | 6.1k | 0.6% |
| System tools | 6.1k | 0.6% |
| Custom agents | 247 | 0.0% |
| Memory files | 385 | 0.0% |
| Skills | 3.4k | 0.3% |
| Messages | 19k | 1.9% |
| Free space | 943.8k | 94.4% |
| Autocompact buffer | 21k | 2.1% |

Delta

| Category | Before | After | Change |
|----------|--------|-------|--------|
| Total | 17.8k (2%) | 35k (3%) | +17.2k |
| Messages | 1.6k | 19k | +17.4k |
| Free space | 961.2k | 943.8k | -17.4k |
| All other categories | — | — | No change |

Expected Behavior

A trivial two-line exchange should consume a small number of tokens proportional to the actual content.

Actual Behavior

~17.4k tokens consumed for a message that should be ~20-30 tokens total. This suggests system reminders, MCP tool lists, or /context command output are being injected into the message history and counted against the Messages category.

Loaded Components

  • MCP tools: 150+ tools (Figma, GitHub, Jira/Atlassian, IntelliJ IDEA, Chrome DevTools)
  • Custom agents: 1 (superpowers:code-reviewer — 247 tokens)
  • Memory files: 1 (CLAUDE.md — 385 tokens)
  • Skills: 17 user skills + 17 plugin skills

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗