It is now Monday, February 2, 2026, early morning.
Resolved the issue of running out of Claude Opus model quota. Found a relay service, and the cost is even lower than DeepSeek. Realized how terrifying the gross profit margins of AI companies can be.
Anthropic's official pricing, calculated per 1M Tokens:
| Model | Base Input Tokens | 5m Cache Writes | 1h Cache Writes | Cache Hits & Refreshes | Output Tokens |
|---|---|---|---|---|---|
| Claude Opus 4.5 | $5 | $6.25 | $10 | $0.50 | $25 |
The relay service provider's quote: Input 0.1890 CNY, Output 0.9470 CNY, also per 1M Tokens.
Calculating at an exchange rate of 1 USD = 7.0 CNY, the official INPUT cost is 35 CNY, while the relay service provider's INPUT cost is 0.189 CNY, a difference of 185.2 times; the official OUTPUT cost is 175 CNY, while the relay service provider's OUTPUT cost is 0.947 CNY, a difference of 184.8 times. In summary, Anthropic's official pricing is approximately 185 times that of the relay service provider.
Clearly, Anthropic's actual cost must be even lower than the relay service provider's price. Even based on the relay service provider's price, Anthropic's gross profit margin is as high as 99.46%. The relay service provider obviously needs to make some profit, meaning Anthropic's gross profit margin is even higher. Moreover, Anthropic certainly does not supply the relay service provider at cost price. It's terrifying to think about.
I used to think AI in this world was expensive; even the "People's AI" (DeepSeek, the king of cost-effectiveness) sells at 2 CNY INPUT + 3 CNY OUTPUT. But a relay service provider can push the cost of a SOTA model (Claude Opus) to less than 1/3 of DeepSeek's price. This world is just too crazy.
Alright, overall, relay servers still carry relatively high privacy risks. However, for many non-sensitive tasks (such as working on open-source projects, learning, research, etc.), using a relay server is still a good choice. After all, the cost difference is just too huge.
