Saving Tokens on Claude Code: Stop Blowing Your Quota

You got a Claude subscription. You blow your limit before the weekend. First reflex: “I need the tier above.”

Wrong reflex. In the vast majority of cases, you’re not consuming too much, you’ve misconfigured Claude Code. Anthropic even admitted in late March 2026 that quotas were “running out too fast,” and part of the problem comes from default settings that cost you dearly without you noticing.

One developer said it bluntly: Claude Code can silently burn 10 to 20 times your token budget. Here are the 7 settings that reverse the trend and easily give you back 50 to 60% of quota.

1. Switch back to Sonnet as default

The most expensive trap. Many people leave Opus as the default model and never change it, even for trivial tasks.

Opus costs roughly 1.7 times more than Sonnet in quota ($5 vs $3 per million tokens on the API side). For the vast majority of work (reading code, simple refactors, generating boilerplate), Sonnet 5 does exactly the same job. Save Opus for genuinely complex problems, architecture, gnarly debugging. Everything else is Sonnet.

Switching back to Sonnet as default recovers up to 40% of your quota with no visible loss of quality.

2. Check that tool search is actually on

This was the number one tip back in 2025: unplug your MCP servers, they eat your context. The GitHub server alone, with its 93 tools, cost 55,000 tokens before your first message.

That’s no longer true by default. Claude Code now ships tool search enabled, and the official docs are explicit: “adding more MCP servers has minimal impact on your context window”. Tool definitions are deferred; only tool names and server instructions load at session start.

Except tool search is off in several configurations: Amazon Bedrock, Google Cloud Agent Platform, Microsoft Foundry, a custom ANTHROPIC_BASE_URL (most corporate proxies), ENABLE_TOOL_SEARCH=false, or a model too old to support it. In those cases you still pay full price.

The reflex: run /context and look at what your MCP servers weigh. If the number is big, you’re in one of the exception cases, and the old advice still applies.

I broke down which servers earn their place here: the best MCP servers for Claude Code.

3. /clear every 45 to 60 minutes

A lingering context gets billed on every message. The longer your session, the more each exchange costs, because the entire history is resent every time.

The rule: /clear every 45 to 60 minutes. Before wiping, ask for a session summary and paste it into the new one. You keep the useful context, you drop the dead weight.

4. Trim your CLAUDE.md

A 2,000-token CLAUDE.md costs 2,000 tokens whether you send 2 messages or 200 in the session. It’s a fixed cost paid on loop.

Keep it short and specific. No essays, no generic “best practices” the model already knows. Only what’s specific to your project and that it can’t guess.

5. Block useless reads with permissions.deny

If Claude Code can read node_modules, build files, dumps, and large assets, it sometimes will. And every useless read is quota spent.

Contrary to what you’ll read all over the place, the .claudeignore file doesn’t exist (it’s still an open feature request). The real mechanism is permissions.deny in .claude/settings.json:

{
  "permissions": {
    "deny": [
      "Read(./node_modules/**)",
      "Read(./dist/**)",
      "Read(./**/*.dump)"
    ]
  }
}

Excluding everything that has no business in the context is passive savings on every session.

6. Hooks, not promises

Writing “never do X” in your CLAUDE.md means paying for an instruction the model can ignore. A hook, on the other hand, actually executes and costs nothing in tokens per turn.

For behaviors you want guaranteed (run a lint, block a file, format), a hook is both more reliable and more economical than a paragraph of instructions.

7. Monitor with /cost and /statusline

You can’t optimize what you don’t measure. The /statusline command permanently displays, at the bottom of your terminal, your cost, your context usage percentage, and session time.

For reference, Anthropic estimates Claude Code’s average cost at around 13 dollars per active developer per day, with 90% of users staying under 30 dollars a day (an estimate it doubled in April 2026, up from 6 dollars). If you’re well above that, it’s not that you work more, it’s that one of these settings is slipping past you.

The bonus trap: the 2026 tokenizer

Worth knowing before switching models: since Opus 4.7, recent models (Opus 4.8, Sonnet 5) ship a new tokenizer that generates roughly 30% more tokens for the same text (up to 35% on code according to third-party benchmarks), at an unchanged sticker price. Your effective bill can climb even though the price didn’t move. Keep it in mind when choosing your default model.

In summary

Before paying 100 dollars more a month:

1. Sonnet as default, Opus in reserve.
2. Tool search on, otherwise MCP servers unplugged.
3. /clear every 45-60 min.
4. Lean CLAUDE.md.
5. permissions.deny to block useless reads.
6. Hooks rather than instructions.
7. /statusline to measure.

The full combo recovers 50 to 60% of quota. Most people who think they need the higher tier actually only need these seven settings.

To go further on config: CLAUDE.md best practices. And to figure out which tier to actually buy: Everyone teaches you to use AI, nobody tells you how to make money with it.