How Claude Code's 1M context burns through limits 3-4x faster

I noticed something unpleasant: the weekly limit on my Max subscription in Claude Code started melting away in about a day. Same projects, same habits — yet a third of the weekly window was gone after 4 hours of work. I figured out what was going on (thanks to Artem Mukhin for the breakdown), and here’s the short version.

What happened

In March 2026, two changes quietly landed:

1M context became the default. Sessions used to hit a wall at 200K; now they stretch to a million.
The “clear context and implement” option was removed from plan mode. The whole long planning conversation now stays in the session.

The UX didn’t change at all. Same response speed, same status bar, /stats shows the same numbers. Meanwhile token usage grew 3.7x in three weeks: from ~27M a day to ~100M.

Why it’s expensive

Claude Code is an iterative session. On every new request, the API receives everything again: the system prompt, CLAUDE.md, tool definitions, and the entire conversation history. That’s billed as cache_read_input_tokens.

1	cache_read = context size × number of messages

You type the same ~2K tokens per message, whether it’s the first or the thirtieth. But by the 30th, the API is re-reading ~250K of accumulated history. With a 200K session you hit the wall after ~50 messages and ran /compact — the cache reset. With a 1M session the conversation runs 300+ messages, context swells to 500K, and every message re-reads the whole tail. Roughly: 5M vs 90M cache_read per “task” — almost an 18x difference.

So why don’t you notice? Because the only way to see it is to crunch your own jsonl files. It’s nowhere in the UI.

The fix — two lines in settings.json

In ~/.claude/settings.json:

{
  "env": {
    "CLAUDE_CODE_DISABLE_1M_CONTEXT": "1"
  },
  "showClearContextOnPlanAccept": true
}

Restart Claude Code and you’re done. The first line brings sessions back to 200K (/compact kicks in earlier, cache reads drop 50-70%). The second restores the “clear context” option in the plan mode menu, so implementation starts from a clean slate without the tail of the discussion.

What else helps stay within budget

/clear between unrelated tasks. The biggest lever. Switched topics — hit /clear. Context is a consumable.
Slim down CLAUDE.md. The file ships with every message: 10K tokens × 300 messages = 3M cache_read a day for nothing. Move rarely needed material into separate files.
Use subagents deliberately. Every subagent gets its own full context from scratch. Fine for parallel search; for sequential tasks it’s simpler to work directly.

Check yourself

Paste this prompt into Claude Code and get your own chart in half a minute:

Read all files in ~/.claude/projects/**/*.jsonl,
deduplicate messages by message.id, extract the fields
timestamp, message.usage.cache_read_input_tokens
and message.model. Group by day for the last
60 days. Show:
1) total tokens per day
2) max cache_read in a single message
3) how many messages per day had context >200K
Find the first day context went over 200K.

If max_ctx jumped above 200K after some date — you’re in 1M mode. It’s not that you started working more: the UX is the same, but the price is 3-4x higher. Two lines of config put everything back.