Three accounts, two months, ~8 billion tokens, ~$5,250 spent. What I actually learned about cache reads, context strategy, and where Claude Code earns its money - pulled straight from the usage reports.

I run Claude Code hard. Two months of pulling daily usage reports across three accounts paint a picture I wish I'd seen on day one - so here it is, with real numbers, not vibes.

The headline numbers

Across one month (April 3 → May 2, 2026) my own account processed ~3.19 billion tokens for ~$2,116 - averaging ~$78/day with one $315 spike on April 18. A second account (similar usage) ran ~$1,928 over the same window. And a March-month run on a third account did 1.88 billion tokens for ~$1,208. Roll it up: roughly 8 billion tokens and $5,250 spent on Claude Code over two months.

If you stop reading here, the only takeaway you need: cache reads dominate everything else.

Cache reads are the cost story

Of those ~3.19 billion tokens on my main account in April, 3.11 billion were cache reads. That's 97%. Input + output combined accounted for under 250 million. The Anthropic prompt cache is what makes Claude Code economically viable for long agentic workloads - without it, the same workload would cost an order of magnitude more.

So the real lever you have is: how warm is your cache? Every clear-context, every long sleep, every cold session pays the input-token tax all over again. Every continuation lets the cache do the heavy lifting at fractional cost.

What spikes the bill

Looking at the daily breakdown, the $200+ days share three patterns:

• Big agentic loops with lots of tool-use - each tool call re-reads context.
• Long sessions that survived multiple compactions - context grows, every read gets bigger.
• Heavy code-editing days - large file reads + writes + lints feed back into the cache cycle.

The cheap days ($1-$15) are the opposite: short scoped sessions, small context, no tool-use loops.

Practical rules I've landed on

1. Stay in cache window. The cache TTL is 5 minutes. If you're sleeping/waiting longer than that, the next wake-up pays for the full context again. Either stay under ~270 seconds OR commit to a much longer wait that amortises the cache miss.
2. Scope sessions tightly. Open a fresh session per logical task instead of pushing one mega-session that accumulates context. Smaller cache footprint per session = lower per-read cost.
3. Use Haiku and Sonnet where you can. The reports show Opus is where the money goes. Drop down a tier for routine edits, lints, status checks - keep Opus for the hard stuff.
4. Watch the cache-read column, not the input/output. That's where 97% of your spend lives.

Is it worth it?

$2K/month for one engineer's tooling is a number that needs justifying. For my work - multi-step agent design, full-stack feature builds, performance audits - Claude Code does the work of probably 1.5-2x my own throughput. At that ratio it pays for itself many times over. But the moment you stop using it for high-leverage tasks and start using it for ergonomic search-and-replace, the ratio flips.

Track your spend. Read your daily reports. The cache numbers will tell you everything you need to know about whether you're using the tool or just paying it.