The Medium headline that stuck — your 10th message costs far more than your first — is the right intuition for Claude usage limits. This guide explains cumulative token waste for chat and Claude Code agents, with sources from Reddit, GitHub, and engineering writeups, plus the practical cuts that actually move the needle.
TL;DR: Claude meters tokens, not chat bubbles. History, tools, and files re-send every turn — so later turns cost far more than early ones. Attack context shape first; then add capacity with clear plan limits.
Myth vs reality
| Myth | Reality |
|---|---|
| “I only sent 10 short prompts” | Each prompt re-includes history + tools + files → tokens compound. |
| “Upgrade fixes everything” | Higher plans add capacity; they do not fix wasteful context shape. |
| “One continue is free” | “Continue” still pays the full conversation tax. |
Companion pillar: How to stop hitting Claude usage limits.
The cumulative context curve
Claude’s meter is tokens. A simplified model of turn N:
cost(N) ≈ system + tools + history(1..N-1) + new_user + new_assistant + tool_resultsThat is why writeups like Ashen Thilakarathna’s Medium guide emphasize later turns dominating. Exact multipliers vary; the shape of the curve is the operational truth.
Agentic coding multiplies the curve
- File reads that stay in history
- Shell / tool JSON payloads
- MCP tool schemas on every turn
- Subagent fan-out (N agents ≈ N context stacks)
Community cost spikes: #41930, #42272.
Seven concrete waste sources
- Infinite threads — prefer Projects / new sessions with a short handoff.
- PDF/blob re-uploads — extract text once.
- Tool soup — disable idle MCP; run
/context. - High thinking on trivial edits — cap
MAX_THINKING_TOKENS. - Lockfiles and node_modules — use
.claudeignore. - Resume of huge sessions — sometimes cheaper to restart (#38029).
- Verbose agent loops — ask for diffs, not full-file rewrites every step.
How Power Claude attacks waste (without relays)
- Token Saver / context optimize before CLI work (local).
- Session recovery so you do not re-explore from zero after a crash.
- official-cli so Anthropic still sees the real client.
- Account groups isolate work/personal memory trees.
Public Power Claude does not market subscription OAuth relays. Prefer official profiles and Anthropic’s own capacity products. See Architecture A vs B.
FAQ
Why does a one-word reply still cost a lot?
Prior turns, tools, and system overhead still travel with the request.
Is the “11× on the 10th message” always true?
Treat it as a teaching metaphor from the Medium article, not a universal constant. Measure with /context and your plan UI.
What is the fastest fix?
New session + Sonnet + disable unused tools, then ignore rules and thinking caps.
Where is the full limits playbook?
How to stop hitting Claude usage limits.
Closing
If you remember one graph: tokens grow with history, tools, and files — not with keystrokes. Attack the curve; then add capacity with clear plan limits.
Independent third-party software. Not affiliated with Anthropic.