Claude doesn't evaluate your messages in isolation. Every time you send a new message, it rereads the entire conversation history from the beginning, which means a chat with 100+ messages can spend 98.5% of its tokens rereading old history and only 1.5% generating the actual response. Message 1 might cost 500 tokens. Message 30 can cost anywhere from 15,000 to 232,000. That compounding, not any single expensive request, is where most AI agent bills actually come from.
Where the invisible overhead comes from
Every turn in Claude Code automatically reloads context you didn't explicitly ask for: your claude.md file, custom system prompts, active files, memory preferences, and every connected MCP server's tool definitions. A single connected MCP server can silently add up to 18,000 tokens of overhead per message. Terminal command output (git status, npm install, test suite runs) gets fed into your context window at full verbosity. And Claude's prompt cache has a strict 5-minute timeout: step away from your desk for six minutes and your next message reprocesses the entire history from scratch at full price.
Systima's benchmark measured the baseline cost of this before you've even typed a prompt: Claude Code sends about 33,000 tokens before reading your first message (OpenCode sends about 7,000), with 27 tool descriptions alone eating roughly 24,000 tokens.
The "token saver" tools don't save what they claim
A wave of open-source tools promises to intercept this overhead. RTK compresses verbose terminal output into dense summaries before it hits Claude's context. Caveman (Kevin Mode) forces terse responses to shrink what gets resent on future turns. Both advertise 60-90% and 65% savings respectively.
JetBrains ran the actual A/B trials, and the marketing numbers didn't hold. RTK advertised 60-90% and increased cost by 7.6% in testing, because its own dashboard measures against a counterfactual maximum that Claude Code's built-in truncation never lets you actually hit. Caveman advertised -65%, measured -8.5%. Ponytail advertised 60-90%, measured -10.3%. Real savings exist. They're nowhere near the marketing.
Cache rebuilds are a separate, quieter drain: a measurement across 185 sessions found 22% of total billing comes from prompt cache rebuilds alone, because a cache rebuild costs 12.5x more than a cache read, and every time a main agent idles behind a subagent past the 5-minute TTL, the next turn re-encodes history you already paid for once.
What actually moves the number
Skip the branded tools and go after the structural causes directly. Never send follow-up correction messages like "no, I meant X." Follow-ups stack permanently onto your history; editing your original message and regenerating replaces the bad exchange instead of multiplying it. Keep claude.md under 200 lines, treating it as a lean index rather than a place to dump full instructions. Reference files surgically with @filename instead of asking Claude to explore a whole codebase. Run /context and /cost regularly to see exactly what's eating tokens, and disconnect MCP servers you're not actively using. Use /clear when switching to an unrelated task, and for long sessions, ask Claude to summarize progress before clearing so you preserve state without carrying the full history forward.
Anthropic also dynamically adjusts how fast your 5-hour rolling session limit drains based on peak usage hours (8am-2pm Eastern on weekdays). Running resource-intensive tasks during off-peak evenings or weekends stretches the same budget further.
The reliability angle you don't see in the pricing
Beyond cost, bloated context degrades output quality through "loss in the middle": as context grows, models pay the most attention to the beginning and end of a session and quietly ignore what's buried in between. Cutting context isn't just a cost optimization. It's also how you keep a long session from getting worse the longer it runs.
FAQ
Why does Claude Code cost more the longer a session runs?
Claude rereads the entire conversation history on every message rather than evaluating each message in isolation, so cost compounds: a message at turn 30 can cost tens of times more tokens than the same message would at turn 1.
Do token-saving skills like RTK actually work?
Independently benchmarked, they save real tokens but far less than advertised. RTK measured a 7.6% cost increase against its advertised 60-90% savings; Caveman measured -8.5% against an advertised -65%.
What's the single highest-leverage habit for cutting token costs?
Never send correction follow-ups ("no, I meant X"). Edit your original message and regenerate instead, since follow-ups stack permanently onto history while an edit replaces the bad exchange.
How much overhead do connected MCP servers add?
A single connected MCP server can add up to 18,000 tokens of overhead per message, since it loads its full tool schema into context on every turn. Disconnecting unused servers is one of the fastest ways to cut baseline cost.
This piece is the companion writeup to the Daily AI Pulse episode "Why Your AI Agent Bill Balloons, And How To Cut It." Watch it on YouTube, or dig into the source links above for the full picture.
More AI breakdowns for solo builders and small teams 👉 joebuildsai.com