Sam Stull

MBA. VP of Sales. I am a bodybuilder. I am a Texas Longhorn fan and a TCU fan.

[Sam Stull @stull73]

On my account a cached token counted about 1/67 as much as a new one against the Claude Code limit.

I'd like to know if that holds on other plans. If you run limitcast results, reply with your numbers.

[Sam Stull @stull73]

Measure your own Claude Code limit ratio:

npm install -g limitcast && limitcast install
Use Claude Code for a day or two, then: limitcast results

It fits your readings to your transcripts. Method: https://t.co/OOLiB2yFBO

[Sam Stull @stull73]

When a Claude Code window is on pace to run out, limitcast Pro's guard stops new subagents. While a window is merely tight, it moves them to Sonnet.

Past 100% it steps aside, so paid extra usage keeps working.

[Sam Stull @stull73]

Weekly Claude limit tip: split what's left of the week evenly across the days until it resets, and watch today's share.

limitcast shows it as "today 6% left", so you find out on Monday, not Thursday.

[Sam Stull @stull73]

When Claude Code limits change, most people find out on Reddit.

limitcast re-measures how many tokens move your 5-hour meter by 1% every day and tells you when that shifts 15% or more. It can ping your phone (ntfy) or Slack too.

https://t.co/SM1xWtbTea

[Sam Stull @stull73]

Trimming 4k tokens of unused skills and MCP servers from my Claude Code starting context saved ~0.3% of a 5-hour window over 1,100 calls.

Real, but small. The big levers are subagent fan-out and cache misses.

[Sam Stull @stull73]

limitcast puts the clock time each Claude Code window runs out at your current pace in your status line:

5h 71% ⚠ out 16:40 · 7d 41% · today 6% left

Local-only, free tier, Codex too.
npm install -g limitcast

[Sam Stull @stull73]

How much did that subagent just cost you?

limitcast 0.11 prints a receipt when each one finishes: "that Explore subagent used ≈2.1% of your 5-hour window (34 calls)".

And you can cap them per window: limitcast budget subagents 15

https://t.co/SM1xWtbTea

[Sam Stull @stull73]

If you pause long enough for the prompt cache to expire, the next call writes your whole context again as new tokens.

In a 200k-token session, that one call is over 1% of the 5-hour limit. (Inferred from my measured new-vs-cached ratio.)

[Sam Stull @stull73]

My Claude Code sessions start at ~100k tokens before I type: system prompt, tools, MCP servers, skills, CLAUDE.md.

Writing that once costs ~0.6% of a 5-hour window. Re-reading it costs ~0.009% per call. Over 300 calls, about 3%.

[Sam Stull @stull73]

Why the first message after a long break hits your Claude Code limit so hard:

The prompt cache expired, so the whole context goes again as new tokens. On my account a cached token counted ~1/67 of a new one.

Numbers and the fix: https://t.co/MS45iXjVnb

[Sam Stull @stull73]

Subagents were 33% of my Claude Code tokens but ~42% of my 5-hour limit.

Each one writes its own context into the cache from scratch. If you keep hitting limits, count your subagents before you trim CLAUDE.md.

[Sam Stull @stull73]

Claude Code shows your usage limit as a percentage. What actually moves it?

On my account, 1% of the 5-hour window was ~166k new tokens, but ~11M cached ones.

Cache reads are cheap. New tokens are not.

[Sam Stull @stull73]

The measurement is built into limitcast, a local CLI for Claude Code and Codex. The free version prints your own ratio after a day or two:
npm install -g limitcast

[Sam Stull @stull73]

Over two days, new tokens were 2.5% of what I sent and ~63% of the limit I used.
Subagents: 33% of tokens, ~42% of the limit.
Trimming CLAUDE.md helps a little. Subagent fan-out and cache misses matter more.

[Sam Stull @stull73]

Method, caveats and numbers:
https://t.co/OOLiB2y7Mg

[Sam Stull @stull73]

How many tokens is 1% of your Claude Code limit?

I fit 61 of my own 5-hour readings against my transcripts:
~166k new tokens, or ~11M cached tokens.
A cache read counted about 1/67 as much as a new token.

[Texas Football @TexasFootball]

Good morning