[Resource] Claude API relay with native Prompt Caching + no prompt storage — cost analysis for CherryStudio users

[lei83314]

Hi all,

I've been running a Claude API relay called Feiyuan API (feiyuanapi.com) for a few months and wanted to share some cost analysis that might be useful for CherryStudio users.

What makes it different from typical reverse-proxy relays:

1. Built on Anthropic's official paid API (no reverse-proxy / no jailbreak)
Feiyuan uses a paid Anthropic commercial account and calls the official API directly. No risk of the "401 account banned mid-workflow" issue that has been hitting many relay services.

2. Prompt/response content not stored
The backend only logs: model name, token count, timestamp. Prompt content and responses are not written to any database.

3. Native Prompt Caching pass-through
The relay passes through Anthropic's cache_control headers without modification, so if you use a fixed system prompt, caching works exactly as documented.

Cost projections (based on Anthropic's published cache pricing — cache reads billed at 10% of input token cost; these are projections, not from a controlled benchmark):

Scenario	Without Cache	With Cache	Est. Saving
5K-token system prompt, 100 calls/day	~500K tokens/month	~50K tokens/month	~89% on cached input

The ~89% figure applies to the cached portion of input tokens only. Actual results depend on your cache hit rate.

Integration with CherryStudio:
Settings → Model Provider → Add OpenAI-compatible provider:

• Base URL: https://feiyuanapi.com/v1
• API Key: your Feiyuan key
• Model: claude-sonnet-4-6 / claude-opus-4-7 / deepseek-chat / qwen3

Models available: Claude Opus 4.7, Sonnet 4.6, Haiku 4.5 + DeepSeek-V3/R1, Qwen3, Kimi

Docs: https://feiyuanapi.com/docs/?utm_source=github&utm_medium=discussion&utm_campaign=feiyuan&utm_content=cherry-studio

Happy to answer questions about caching setup.
Telegram group: https://t.me/feiyuanapi_group