If you starred TokenTamer waiting for the big feature — this is it.
What's New: Long-Lived Session Hijacking
We just shipped Anthropic prompt caching exploitation — the single most powerful cost-cutting feature in TokenTamer.
The Numbers
|
Before |
After |
| Cached input cost |
— |
$0.30 / Mtoken |
| Regular input cost |
$3.00 / Mtoken |
$3.00 / Mtoken |
| 50-turn Claude Code session |
~$5.00 |
~$0.50 |
Same model. Same reasoning. Same output. Just smarter billing.
How It Works
Claude Code (and most agents) re-send the entire conversation every turn. That means turn 50 sends messages 1-50 all over again. Expensive.
TokenTamer now injects cache_control breakpoints at three stable positions:
- After the tools array (rarely changes)
- After the system prompt (fixed for the session)
- After the conversation prefix (everything except the last 2 turns)
Anthropic caches the prefix. You only pay for the new tail. On a long session, 80-90% of your input tokens become cached — and cached is 90% cheaper.
Verification
Run TokenTamer, use Claude Code, then check the response headers:
curl -I http://127.0.0.1:8000/v1/messages ...
# X-TokenTamer-Cache-Breakpoints: 3
# X-TokenTamer-Cache-Tokens: 12400
If you starred TokenTamer waiting for the big feature — this is it.
What's New: Long-Lived Session Hijacking
We just shipped Anthropic prompt caching exploitation — the single most powerful cost-cutting feature in TokenTamer.
The Numbers
Same model. Same reasoning. Same output. Just smarter billing.
How It Works
Claude Code (and most agents) re-send the entire conversation every turn. That means turn 50 sends messages 1-50 all over again. Expensive.
TokenTamer now injects
cache_controlbreakpoints at three stable positions:Anthropic caches the prefix. You only pay for the new tail. On a long session, 80-90% of your input tokens become cached — and cached is 90% cheaper.
Verification
Run TokenTamer, use Claude Code, then check the response headers: