Per turn token cost: the six line items every Claude Code request pays for

Most "why is Claude Code expensive" posts stop at one line: CLAUDE.md fires every turn. That is true, and it is one of six things that fire every turn. The bill on any given API call is six concatenated billables, all priced at the input rate, plus the output. If you want to lower the per-turn number, you need to know which of the six you can actually touch.

This page walks the math line by line, with the analyzer formula from src/lib/analyzer.ts as the spine and the current Anthropic pricing as the rate sheet.

1. The six line items in one turn

A Claude Code turn is one POST to /v1/messages. The Messages API is stateless, so the harness rebuilds the entire context on every call. Six things land in that payload, all charged at the input rate, before the model writes a single output token:

1. Anthropic's system prompt. About 2,800 tokens. Hidden, not editable, charged at input rate. Includes safety, tool-use scaffolding, and the model's built-in identity.
2. Your CLAUDE.md. Project CLAUDE.md + every nested CLAUDE.md the harness walked up to + the user-global ~/.claude/CLAUDE.md. All concatenated, all firing every turn for the session.
3. Tool definitions. Schemas and descriptions for Bash, Edit, Read, Glob, Grep, Write, plus any MCP servers wired in. Bash alone adds 245 tokens; a verbose MCP server can add 1,500. This is usually the easiest line item to cut.
4. Tool-use system prompt. Fixed cost when tool_choice is auto or none: 346 tokens on every Claude 4.x model. 313 tokens when tool_choice is any or a specific tool.
5. Conversation history. Every previous user message, assistant message, tool call, and tool result from this session. Grows linearly with turn count until a /compact or a new session.
6. Your new user message + attached files. The text you typed plus any files, screenshots, or pasted content the harness inlines. Counts toward input tokens.

Numbers above are for a mid-session turn after some history has accumulated. The first turn of a session pays for items 1-4 plus a tiny history; turn 30 pays for items 1-4 plus 30 turns of accumulated history. The growth is in line item 5.

2. The analyzer's formula, from the source

Our analyzer at ccmd.dev gives you the CLAUDE.md slice of one turn (line item 2). It does not know about lines 1, 3, 4, 5, or 6 because those depend on your harness configuration and your session state. The CLAUDE.md slice is the part you can score from a static file, and it is the part that fires unconditionally on every turn:

The hardcoded OPUS_IN_PER_M = 15 was Opus 4.1's input rate. Anthropic dropped Opus 4.5, 4.6, and 4.7 to $5/M, so the analyzer overstates input cost by 3x on the current default Claude Code model. Output rate also dropped: $75/M on 4.1 down to $25/M on 4.7. The ratio is preserved; the absolute is not. Divide the analyzer's reported cost by 3 if you're on Opus 4.7.

3. Rates by model (current vs. deprecated)

Same shape of turn, four different bills depending on the model. Numbers are from the Anthropic pricing docs, verified 2026-05-18. The deprecated column is what you still pay if your harness is pinned to Opus 4.1 or earlier.

The 3x gap between Opus 4.7 and Opus 4.1 is the single biggest per-turn lever most teams have not pulled. If your CLI is still pointing at claude-opus-4-1, every turn pays the old rate, your bill is 3x higher than it should be, and the tokenizer change in 4.7 (which can use up to 35% more tokens for the same text) does not even kick in to offset it. The default model in current Claude Code releases is 4.7 unless you override.

4. Where prompt cache changes the per-turn bill

Anthropic's prompt cache reads cost 10% of the base input rate. On Opus 4.7 that is $0.50/M instead of $5/M. The catch: the cached prefix has to be byte-identical between requests, within a 5-minute (or 1-hour) window. Line items 1, 2, 3, and 4 are usually stable enough to cache; line items 5 and 6 are fresh on every turn.

The single biggest cache killer is a volatile string in the first 20 lines of CLAUDE.md. The analyzer flags this as a cache_bust finding:

One ISO date, one today is, one this session in the first 20 lines, and the entire prefix re-pays full input cost on every turn for the rest of the session. The bytes flagged are 30 or 40; the cost is the whole file.

5. Three levers to cut the per-turn number

Of the six line items, four are touchable. The other two (Anthropic's system prompt and Anthropic's rate sheet) are fixed. In order of impact per minute of work:

1. Confirm you're on Opus 4.7, not 4.1. Check your harness config or run claude --version. If the default model id is claude-opus-4-1 or earlier, switching to claude-opus-4-7 cuts input cost by 3x and output by 3x. This is the single highest-leverage change on this page.
2. Disable MCP servers you don't use. Each wired MCP server adds 200 to 1,500 tokens of tool definitions to line item 3. A typical user-global MCP config has 5-10 servers loaded, of which 1-2 actually get invoked in any given session. Comment out the rest in ~/.claude.json or your project config. Saves more per turn than rewriting CLAUDE.md.
3. Score and trim CLAUDE.md. Paste your file into ccmd.dev. Fix any cache_bust findings first (one line, 10x bill recovery). Then attack bloat, duplicate, and vague findings in order of tokenSavings. For each cut, the analyzer recommends an installable skill from skillhu.bz that replaces the lines without firing every turn.
4. Run /compact before line item 5 gets out of hand. Conversation history is the only line item that grows linearly without an upper bound. A 100-turn session can have 80,000 tokens of history alone. The /compact command rewrites history into a shorter summary; the next turn pays for the summary instead of the full transcript. Sometimes saves more per turn than anything else if you forgot to compact for an hour.

6. Related reads

• Do CLAUDE.md rules fire on every turn? — proof from the analyzer source plus the four firing surfaces (CLAUDE.md, skills, hooks, MCP) and which kinds of rules belong where.
• CLAUDE.md token waste per turn — the seven categories of bytes that fire every turn and produce zero behavioral signal, ranked by dollar impact.
• CLAUDE.md token cost audit — what a per-line audit actually looks like on a real 6,000-token file.
• Layered CLAUDE.md token cost — what happens when the harness walks up and concatenates three CLAUDE.md files before turn one.