Cut Your Claude Code Bill

Claude Code is often a developer's biggest AI bill. Every step of an agentic session reads files, runs tools and resends a growing context, and it's easy to leave a heavy model working on light tasks. This guide lowers the cost in three ways: a lower price per token, the right model for each job, and hard limits so a long session can't surprise you.

1. Create a dedicated key with a daily limit

On the API Keys page of your dashboard, create a key just for Claude Code and name it, for example, "claude-code laptop". Then set a daily spend limit that fits how you work; $10 a day is a sensible starting point for one developer.

2. Connect Claude Code to GreenTokens

Claude Code reads its connection settings from environment variables. Put them in its settings file so they apply to every session:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.greentokens.io",
    "ANTHROPIC_AUTH_TOKEN": "sk-gt-…"
  }
}

Use the base URL exactly as shown, without /v1. Start Claude Code and send a short message; it should appear in your dashboard's Requests log within seconds. The Claude Code docs page has the same setup as shell variables.

3. Map each model slot to the right model

Claude Code uses three model slots: Opus and Sonnet for your conversation, and Haiku for small background tasks. Mapping them explicitly controls which model, and which price, each one uses:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.greentokens.io",
    "ANTHROPIC_AUTH_TOKEN": "sk-gt-…",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-5-5",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-5",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-haiku-4-5"
  }
}

The Haiku slot handles frequent small jobs, so keeping it on the cheapest model saves money on every session without affecting your main conversation. Prices for every model are on the models page.

4. Pick the model per task

Switch models inside a session with /model. A simple habit makes a large difference:

• Use Sonnet for everyday work: edits, tests, refactors, explanations. It costs half of Opus per token.
• Switch to Opus for the hard problems: tricky debugging, large design decisions. Switch back when they're done.
• Don't leave Opus selected for a whole day of routine changes.

5. Keep the context lean

Every step resends the session's context, so a bloated context costs you on every turn. Two built-in commands help:

• /clear starts fresh. Use it when you move to an unrelated task, instead of carrying the old conversation along.
• /compact summarises the conversation so far into a shorter form. Use it when a long task is still in progress but the history has grown.

Also avoid asking Claude Code to read large generated files, lockfiles or build output unless the task needs them. Everything it reads becomes part of the context. What a context window is explains why this adds up.

6. Check what it costs

After a day of normal use, open the Requests page and filter by your Claude Code key. You'll see every call with its model, tokens and cost, and the totals for the period.

• If most spend is on Opus, check whether those tasks needed it.
• If you hit the daily limit on normal days, raise it to match your real usage.
• If you never get near it, lower it: a tight limit is a better safety net.

Next steps

• Building your own agent too? Add prompt caching to it.
• Use the same key setup in other tools: Codex CLI, opencode and more in the docs.