Summary
Since approximately March 28-29, 2026, my token consumption on the Max plan ($100/month) has increased dramatically — roughly 3-5x — without any changes to my setup.
Before (working fine for months)
- 5+ hour coding sessions possible
- Multiple terminals in parallel
- 30-60 minute cooldown between heavy sessions
- Same plugins, same model, same configuration
Now (since ~March 28-29)
- 1-2 hours in a single terminal → limit reached
- ~42% of the 5-hour rolling window consumed by a moderate coding session
- Frequent rate limit warnings
Setup (unchanged for months)
- Plan: Max ($100/month)
- Model: Opus 4.6 (1M context) —
claude-opus-4-6[1m]
- Claude Code version: 2.1.87
- OS: Debian 12 (remote server via SSH)
- Plugins: 57 enabled (unchanged)
- MCP Servers: 9 configured (unchanged)
- Settings:
alwaysThinkingEnabled: true, effortLevel: "high", CLAUDE_CODE_MAX_OUTPUT_TOKENS: 32000
Token Analysis (estimated for today's session)
A single coding session (building a maintenance page — ~1000 lines of new code, 4 deploys, 15 back-and-forth messages) consumed approximately 42% of the 5-hour limit.
Estimated breakdown:
| Category |
Estimated Tokens |
| System context overhead per request |
~40,000 |
| Context overhead × ~30 turns |
~1,200,000 |
| Conversation history (growing) |
~600,000 |
| Tool results (file reads, bash output) |
~300,000 |
| Output (responses + code) |
~150,000 |
| Total |
~2,250,000 |
The per-request overhead (~40k tokens) is dominated by:
- Plugin/skill descriptions (57 plugins → ~15k tokens)
- MCP server instructions (~5k tokens)
- Project rules + CLAUDE.md files (~12k tokens)
This overhead existed before and was not a problem. The same configuration worked fine for months.
What changed?
Nothing on my end. The plugin list, MCP servers, rules files, model selection, and Claude Code settings have been identical for weeks/months. The spike in consumption started around March 28-29.
Possible server-side causes:
- Change in token accounting/billing for the 1M context variant?
- Change in rate limit calculation for Max plan?
- Change in how system prompts / plugin descriptions are counted?
- Claude Code update changing context management?
Expected behavior
Token consumption consistent with the same setup 1 week ago — i.e., 5+ hour sessions without hitting limits after 1-2 hours.
Environment
Claude Code: 2.1.87
Model: claude-opus-4-6[1m]
Plan: Max ($100/month)
OS: Debian 12
Summary
Since approximately March 28-29, 2026, my token consumption on the Max plan ($100/month) has increased dramatically — roughly 3-5x — without any changes to my setup.
Before (working fine for months)
Now (since ~March 28-29)
Setup (unchanged for months)
claude-opus-4-6[1m]alwaysThinkingEnabled: true,effortLevel: "high",CLAUDE_CODE_MAX_OUTPUT_TOKENS: 32000Token Analysis (estimated for today's session)
A single coding session (building a maintenance page — ~1000 lines of new code, 4 deploys, 15 back-and-forth messages) consumed approximately 42% of the 5-hour limit.
Estimated breakdown:
The per-request overhead (~40k tokens) is dominated by:
This overhead existed before and was not a problem. The same configuration worked fine for months.
What changed?
Nothing on my end. The plugin list, MCP servers, rules files, model selection, and Claude Code settings have been identical for weeks/months. The spike in consumption started around March 28-29.
Possible server-side causes:
Expected behavior
Token consumption consistent with the same setup 1 week ago — i.e., 5+ hour sessions without hitting limits after 1-2 hours.
Environment