This is methodologically flawed, as bytes only weakly correlate with tokens.
Unless you're sending identical requests, you can't expect the same token counts for any given of bytes, or that a slightly longer (but different) message will lead to more tokens than a slightly shorter one, or vice versa.
> The numbers came from the same project and the same prompt across versions.
I'm pretty sure the tester checked. If the request format is the same (which it is, given it uses the same as Anthropic's stable public API) and the same prompt/messages then bytes will correlate pretty well.
I had the same suspicion so made this to examine where my tokens went.
Claude code caches a big chunk of context (all messages of current session). While a lot of data is going through network, in ccaudit itself, 98% is context is from cache.
Granted, to view the actual system prompt used by claude, one can only inspect network request. Otherwise best guess is token use in first exchange with Claude.
On the free plan, I hit the limit instantly by uploading one 45kb PDF and one prompt. Even for a free plan, I expect a bit more. Oh well, local models can be pushed to do what I need.
This is methodologically flawed, as bytes only weakly correlate with tokens.
Unless you're sending identical requests, you can't expect the same token counts for any given of bytes, or that a slightly longer (but different) message will lead to more tokens than a slightly shorter one, or vice versa.
> The numbers came from the same project and the same prompt across versions.
I'm pretty sure the tester checked. If the request format is the same (which it is, given it uses the same as Anthropic's stable public API) and the same prompt/messages then bytes will correlate pretty well.
The prompt may be the same, but the project context would have have surely changed. User prompt itself is unlikely to be ~200KB.
According to the linked repo, they add the cost of those internal prompts to your bill.
I had the same suspicion so made this to examine where my tokens went.
Claude code caches a big chunk of context (all messages of current session). While a lot of data is going through network, in ccaudit itself, 98% is context is from cache.
Granted, to view the actual system prompt used by claude, one can only inspect network request. Otherwise best guess is token use in first exchange with Claude.
https://github.com/kmcheung12/ccaudit
I got kinda obsessed with observability a month ago and wired together a full stack for personal use.
https://github.com/simple10/agent-super-spy - llm proxy + http MiTM proxy + LLMetry + other goodies
https://github.com/simple10/agents-observe - fancier claude hooks dashboard
It started as a need to keep an eye on OpenClaw but is incredibly useful for really understanding any agent harness at the raw LLM request level.
On the free plan, I hit the limit instantly by uploading one 45kb PDF and one prompt. Even for a free plan, I expect a bit more. Oh well, local models can be pushed to do what I need.
I don’t buy it. The same problem was reported in Claude.ai at the same time which means same underlying root cause.
What is the system prompt for $1000 Alex (RIP)?
a git status + "recent commits" (that will _always_ change) https://github.com/anthropics/claude-code/issues/47107
and a missing cache-mark that will make skills & project-claude.md cachemiss every time too https://github.com/anthropics/claude-code/issues/47098