成本对比：$20 订阅下 Opus 5.5 每 token 比 V4.1 Flash 便宜 1.6 倍
作者算账称同样 $20，Claude Pro 约 29 亿 token、DeepSeek V4.1 Flash 约 18 亿，因 96% 计费来自缓存；但按 API 标价同样 token 是 $1650 对 $33。
This cannot be true. Claude Opus 5.5 is 1.6x cheaper per token than DeepSeek V4.1. Same workload, same $20. $20 of Claude Pro: ~2.9 billion tokens. $20 of DeepSeek V4.1 Flash (off-peak): ~1.8 billion tokens. $0.0069 per million tokens vs $0.0113. The lab with the higher API price sells the cheaper token. At API list price, those same 2.9 billion tokens cost ~$1,650 on Opus 5.5 and ~$33 on V4.1 Flash. 50x the other direction. Both numbers are true, and for the same reason on both sides: ~96% of what gets counted is cache. On Claude that is cache reads, the agent re-reading its own context at 5% of the input price. On DeepSeek that is cache hits at $0.003/M. DeepSeek at least publishes the shape of it. Its DeepInfer release: 155 trillion tokens a day on one cluster, 96% KV-cache hit rate, ~$38 billed per GPU-hour at off-peak prices. DeepSeek sells the same work 50x cheaper on paper and 1.6x more expensive in practice. What is the correct pricing model here?