I use Claude Code every day to build a SaaS product. This week, the usage meter hit 95% before Wednesday. I assumed I had somehow written a mountain of code.
I hadn't.
Every request is logged locally, so I added it up. Here's what three days of work looked like (October 12–15):
- 65,000 requests
- 12.6 billion tokens processed
- 14 million tokens of actual output
That means 99.9% of my tokens weren't code. They were the model re-reading what it already knew.
The Real Cost Is Context, Not Code
97.6% of all tokens were cache reads. Every turn, the agent resends the entire conversation so far: files read, tool results, earlier decisions.
When the context sits at 200k tokens, a one-line reply still costs 200k tokens of reading.
Two thirds of my volume came from requests with more than 200k tokens of context. Only 2.5% came from requests under 50k.
A long session doesn't cost more because it does more. It costs more because every new turn is more expensive than the last.
Subagents Are a Multiplier
My workflow splits work into small tasks. For each task, one subagent implements and two review — one for spec conformance, one for code quality. It produces clean results.
It also produced 78% of my requests and 53% of my tokens.
Each subagent starts with a fresh but still large context, runs many tool calls, and each call re-reads that context. One task becomes dozens of requests.
Five Sessions Were 75% of the Week
Out of 46 sessions, five accounted for three quarters of the total volume. These were multi-hour runs where the context kept growing and I never started over. One of them averaged almost 400k tokens of context per request.
The rest of the breakdown:
| Model | Share of Tokens |
|---|---|
| Sonnet | 82% |
| Opus | 14% |
| Fable (premium tier) | 3.6% — almost all in a single day |
What I'm Changing
- One session per block of tasks. I'll clear or compact well before the context reaches 150k.
- Fewer parallel subagents. One review cycle for routine work instead of two.
- A smaller model for routine subagents like review and exploration.
- Premium models only for planning and hard decisions.
- A small local script that sums my logs every morning, so I see the curve before the meter does.
A Caveat on the Numbers
These are raw token counts. Cache reads are not billed or limited the same way as fresh tokens, so this explains the volume, not the exact percentage shown on the usage meter. The pattern is what matters.
The Takeaway
If your AI coding quota disappears faster than your output suggests, look at your context size and your subagent fan-out before you blame the model.
Tokens aren't spent on the answer. They're spent on remembering the question.
How are you tracking your agent usage? Do you reset sessions deliberately, or only when something breaks?
