我追踪了每个 API 调用 30 天。60% 在错误的模型上浪费了钱。

Tracking API calls over a month revealed that out of approximately 3000 calls, 1800 (60%) were simple tasks such as file reads, grep, and basic Q&A; 750 (25%) were medium complexity tasks like code refactoring, test generation, and summarization; and 450 (15%) were complex tasks involving architecture decisions and multi-file debugging. All calls were directed to Claude Sonnet at a cost of $15 per million tokens. The simple tasks could run on DeepSeek-V3 for $1.80 per million tokens, while medium tasks performed adequately on GPT-4o at $5 per million tokens. Only the complex tasks genuinely required Sonnet. This routing adjustment significantly reduced monthly costs from around $240.