Why did this prompt cost so much?

This started over Slack. Someone mentioned that one of their Claude Code sessions seemed to use more tokens than other sessions. Nobody in the thread could say why. We all had guesses: Context? Big files? Too many tool calls? The model doing something weird in the background?

One(-ish) prompt(s) later, and we had a Claude Code plugin (code name TokenTab) that showed tool breakdowns, background calls, and caching.

Stuff we learned by staring at it

The big one is context. Every call re-reads the whole conversation. The longer the session, the more context. Makes sense. What we didn't realize was the cost of coming back to old sessions!

Some other things that surprised us:

• Going to get coffee can cost you. The prompt cache lasts an hour. Come back after it's expired, and your next message rewrites the entire context at the expensive rate. TokenTab now catches that one before it's sent and says, "This'll cost $X; a fresh session would cost $Y, your call."
• Those "suggested next prompt" things cost money. After each turn, Claude Code makes a call to suggest what you might type next, and that call re-reads your context too. $0.08 in a short chat, $0.20 in a long one.
• Cheaper models aren't always cheaper. We had a Haiku helper on a browser form job make 4x the calls Opus did. 
  

Then we gave it a budget

We're Supertab. We think about small payments adding up for a living, so once we could see the spend, we obviously had to try putting a limit on it.

We tried a few ways and ended up with two that felt right: a session budget and a daily budget.

Here's the interesting bit. It doesn't tell Claude how to save money. It just shows the numbers. But with the numbers in front of it, Claude started behaving differently:

• Claude told me to run a command myself. Instead of running a terminal command and handling all the output, it just handed me the command and said, Running this is quicker and easier if you just do it.
• Claude stopped writing essays. No list of every edge case and what-ifs. Just the thing I needed to finish what I was doing.
  

Try it

To be clear, this is a free, open-source beta. No account, no sign-up, and nothing leaves your machine: it reads what Claude Code already knows and writes a log to a folder on your disk.

Please note: It only works on Claude Code (not chat or Cowork).

/plugin install tokentab --marketplace getsupertab/claude-marketplace

Start a new session (or /reload-plugins) and it's counting from your first prompt. A pane opens beside the chat with the current session's cost broken down by tool and by type. It also lists the cost of all sessions in the current month. 

You'll need Claude Code 2.1.304+, macOS and Python 3.

Auto-update can be enabled in the terminal via: /plugin → Marketplaces → supertab → Enable auto-update.

The Desktop app turns auto-update off, so there you'll run claude plugin marketplace update supertab now and then.

Why are we sharing this?

This was meant to be a one-off test, but the changes to how the agent operated seemed like enough to share it with others to see if there is a future product here.

Give it a try and let us know:

• Budgets: do you care? Would you actually set one? What number? Or is just seeing the spend enough?
• Your tricks for keeping sessions cheap. We'll collect the good ones and share them.
• Anything else, good or bad.

Email us at product@supertab.co. And if you want to know more about what Supertab does, check us out at supertab.co.

‍