Claude Code token cost: what a session actually burns

Quick answer: expect dollars per session, not cents.

Claude Code is an agentic tool, not a chat window. One request from you can trigger dozens of model calls, and each call re-sends the entire growing conversation as input tokens. That makes input volume, not output, the driver of cost. Output tokens are usually a rounding error next to input. Approximate community-reported numbers put a serious hour-long session on a medium repo at roughly 5 to 20 million tokens. At Anthropic direct rates for Sonnet 4.6 ($3 per million input, $15 per million output), most sessions land somewhere between about $2 and $40 before caching discounts, and heavy Opus days go well past that. Prompt caching pulls real bills below the naive math, but the shape holds: long sessions on big repos cost real money. You have three levers. Reduce what enters the context window, route routine work to cheaper models, and lower the per-token rate by changing where requests are billed. The rest of this page works through each one with concrete numbers.

Where the tokens actually go.

Two mechanics soften this. First, Anthropic prompt caching: Claude Code marks stable prefixes as cacheable, cache writes cost a premium, and cache reads bill at about one tenth of the normal input rate. Second, context compaction: the /compact command summarizes history and drops the raw transcript. Both help a lot, and neither changes the fundamental rule: everything sitting in the window is paid for on every call that includes it. This is why the same prompt costs wildly different amounts in different repos. A question asked in a fresh session with a lean CLAUDE.md is cheap. The same question asked 30 turns into a session that has read a dozen large files rides on top of a six-figure token window.

• System prompt and tool definitions: roughly 15 to 20K tokens, re-sent with every single model call.
• CLAUDE.md, rules files, and MCP tool schemas: loaded at session start and carried in every subsequent call.
• File reads: a 500-line source file is roughly 5 to 7K tokens, and it stays in the window after being read once.
• Tool results: test output, grep hits, and terminal logs get appended to the conversation and re-billed as input on every later turn.
• The agentic loop itself: a single "fix this failing test" can produce 10 to 40 model calls before Claude Code reports back.

Realistic session cost math.

Model it turn by turn instead of guessing. A typical mid-size session runs about 40 model calls. Once a few files and some test output have accumulated, the full window averages around 90K tokens per call. Forty calls at 90K is 3.6M input tokens. Output stays small: 40 calls at roughly 1,000 tokens each is about 40K output tokens. The table below prices three session profiles at Anthropic direct list rates. Treat these as ceilings: prompt caching discounts the repeated prefix heavily, so cached sessions usually come in lower. The pattern still matches what developers report. Occasional afternoons cost a few dollars; running Opus as a daily driver on the API can clear several hundred dollars a month.

Plan vs API: which way is Claude Code cheaper?

Anthropic sells Claude Code two ways: bundled into flat-rate subscription plans with usage caps enforced in rolling windows, or metered per token on the API. Claude Code API pricing is simply the model rates above applied to whatever your sessions burn, with no cap and no cut-off. The break-even is plain arithmetic. If your typical session costs about $11 on Sonnet at direct rates, a $20 plan pays for itself at roughly two sessions a month, provided you stay under its caps. The catch is the caps themselves: heavy agentic use is exactly the pattern that trips rolling limits, and a flat rate stops being cheap the moment you are locked out mid-session. API billing inverts the trade. You are never cut off, and you pay for exactly what you burn. Many heavy users end up hybrid: a plan for interactive work, an API key for CI jobs and overflow.

How to cut Claude Code token cost.

The last two levers compound. Routing a cheap model through a discounted endpoint can turn a $10 session into well under a dollar for routine work. One gateway option is APIsRouter, an OpenAI-compatible API with pay-as-you-go billing and no subscription: global models are priced 20% below official list, Chinese models sit below their official rates, the first top-up adds +100% balance, and the no-signup checkout at /topup takes payment first and emails the key. The catalog below shows what a routing ladder looks like in practice. Kimi K2.6 (kimi-k2.6) is also available if you prefer it as the budget tier.

• Keep CLAUDE.md lean. Every line is a recurring charge on every model call in every session.
• Start a fresh session per task and run /compact when a long session drifts. A 150K-token window bills 150K on each call.
• Scope reads. Point Claude Code at specific files instead of letting it crawl, and exclude build output and vendored directories.
• Disable MCP servers you are not using. Their tool schemas ride along in the system prompt on every call.
• Route by model. Keep Opus for architecture and gnarly debugging, run Sonnet as the default, and push boilerplate, test scaffolding, and commit messages to cheaper models.
• Lower the per-token rate itself by pointing the base URL at a gateway instead of billing at direct list prices.

Config example: repoint Claude Code and verify.

Claude Code reads ANTHROPIC_BASE_URL on startup, so switching where tokens are billed is a settings change, not a workflow change. Put the endpoint and key in ~/.claude/settings.json and every session picks them up. Model IDs stay the same, tool execution stays local on your machine, and rollback is commenting out one line. For OpenAI-compatible tools and scripts, the base URL ends at /v1 (https://api.apisrouter.com/v1). Verify the key with a one-off curl before pointing your daily driver at it:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.apisrouter.com",
    "ANTHROPIC_AUTH_TOKEN": "sk-APIsRouter-...",
    "ANTHROPIC_MODEL": "claude-sonnet-4-6",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-4-7",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-4-6"
  }
}

FAQ