A consultancy firm tested the '80x cheaper' claim regarding DeepSeek by renting four Nvidia H200 GPUs. The setup cost $18.37 per hour on-demand, totaling approximately $13,200 for a standard month. This figure is more than double the firm's September bill for Anthropic's Claude Code, which amounted to about $5,500.
During the test period from September 1 to 27, the server processed 2.03 million model calls and 388.5 billion tokens. Of these, 374.2 billion tokens were cached, while 393 million were written. The firm estimated that the same volume of tokens on the DeepSeek API would cost between $3,500 and $7,000, compared to approximately $140,000 for Claude Opus 5.5.
Technical issues prevented the model from writing code, as reviewers identified security risks that kept the agents offline. Consequently, the system operated as 48 to 64 read-only reviewer agents, which processed 2,377 folders and filed 32 bug reports. The hardware, rented at a spot rate of $9.19 per hour, was reclaimed by the provider shortly after testing concluded.
