Call Center Firm's H200 Test Undercuts DeepSeek's Cost-Advantage Claims
A consultancy rented high-end Nvidia hardware to benchmark DeepSeek against Claude, only to find that using the model's own API remained cheaper than self-hosting.
The Call Center Doctors, which operates and manages call center operations for client businesses, leased a quad-H200 system on September 27 to run DeepSeek V4.1 Flash as a replacement for Claude Opus 5.5 in its Code agents. The experiment, detailed in a published assessment, was designed to test whether DeepSeek truly delivers the "80x cheaper" performance that has circulated in industry discussions.
The rented hardware configuration achieved approximately 213 tokens per second when processing the firm's actual coding workload. At on-demand pricing, this setup cost $440.88 daily—substantially higher than the $184–$223 per day the same work would cost through DeepSeek's API service. The consultancy ultimately returned to Opus 5.5, which proved more economical than operating the rented hardware at standard rates.
DeepSeek's API charges $0.15 per 1 million new input tokens during off-peak hours, $0.003 for cached input tokens, and $0.60 per 1 million output tokens, with prices doubling during peak periods. Anthropic's Claude Opus 5.5 carries a steeper price tag: $4 per 1 million input tokens and $20 per 1 million output tokens, with cached reads at $0.20 per 1 million tokens. The Call Center Doctors accessed Claude through subscription arrangements rather than the API.