If you're using Claude Code daily, you've probably noticed: most of your API costs come from re-sending the same data over and over. System prompts, tool schemas, conversation history — 85-95% of tokens per turn are identical to the previous request.
Anthropic's prompt caching helps, but with a critical limitation: the cache TTL is only 5 minutes. Take a coffee break during a coding session, and your cache is gone.
Synrouter takes a different approach: session-lifetime caching.
How it works
Instead of tying cache lifetime to a fixed clock, Synrouter binds it to your agent session:
- First turn of a session — full request is sent, cache is written
- Subsequent turns — cache is checked and reused
- Session ends — cache is evicted
No 5-minute timer. No cold starts after a break. Your 2-hour coding session stays cached for the full 2 hours.
The numbers
For a typical Claude Code session (50 turns, ~50K tokens/turn):
| Scenario | Cost per 100 turns | |---|---| | No caching (baseline) | $75.00 | | Anthropic prompt caching (85% hit rate, 5min TTL) | ~$35.00 | | Synrouter session cache (85% hit rate) | $17.60 |
That's a 76% reduction vs baseline, and 50% less than Anthropic's built-in caching.
Drop-in replacement
The best part: you don't need to change your agent code.
One line. That's it. All caching, compression, and optimization happens transparently server-side.
What about other agents?
Synrouter works with any OpenAI-compatible agent:
- Claude Code — Anthropic SDK via base_url swap
- Codex CLI — OpenAI-compatible endpoint
- Cursor Agent — Full compatibility
- Factory Droid — Drop-in replacement
- Hermes / OpenClaw — General agent support with extended tool routing
Get started
Synrouter is in Early Access. Sign up for the trial to get your API key and start cutting costs today.
This is the first post in our series on agent inference optimization. Next up: "Session-Aware Caching for Cursor & Codex."