A consultancy rented four Nvidia H200 servers to test the claim that DeepSeek is 80 times cheaper than Claude, and found the math depends heavily on how the comparison is framed. The firm ran DeepSeek as 48 to 64 read-only reviewer agents because its reviewers kept finding ways agent code could break out of its sandbox, including a settings file in a shared temp folder that could let agent code run as an administrator. That flaw kept the code-writing agents offline, so DeepSeek never wrote actual code during the test. The rented box was returned by the provider minutes after the last test finished.
The DeepSeek API lists $0.15 per 1 million new input tokens off-peak, $0.003 for cached input tokens, and $0.60 per 1 million output tokens, with peak prices double those rates. Anthropic's Claude Opus 5.5 is listed at $4 input and $20 output per 1 million tokens, with cache reads at $0.20. The consultancy used Claude through subscriptions rather than the API.
Running the model stably required five attempts, each with 10 to 15 minutes of loading. The box bills whether it is busy or not, and at the $18.37 hourly on-demand rate a standard-length month comes to about $13,200, more than twice the consultancy's September Claude bill. One box can process roughly 20 billion tokens a day by the consultancy's math, most of it re-reading, against the 51 billion tokens its agents used on their busiest day. Coding agents mostly resend conversations, and 96 percent of what they provided the model was stale text, according to the consultancy's logs, with about 1,000 old tokens consumed for every new token written.
One server running from Sept. 1 to 27 logged 2.03 million model calls, 388.5 billion tokens read, 374.2 billion of them cached, and 393 million written, producing 5,610 merged changes. The consultancy's September Claude Code subscriptions totaled about $5,500. The same 27 days of tokens on DeepSeek's API would cost between $3,500 and $7,000 depending on peak pricing, with an estimated average near $4,200 if usage were spread evenly across the week. Against Claude Opus 5.5 list prices, the same tokens would come to about $140,000, more than 25 times the subscription cost, and that flat-rate discount is where the 80x figure originated.
The consultancy also priced each merged code change at about $1 on Claude and about $0.63 on DeepSeek's API, assuming DeepSeek did the same work with the same tokens. Its estimate of $1.15 to $4.90 for DeepSeek assumes the model would need more tokens, succeed less often, and require Claude to check its work. The rented box ran at a $9.19-an-hour spot rate. Based on the V4 generation, in which Pro outscored Flash, better performance can be expected from DeepSeek V4.1-Pro, which has no firm release date.
More software news from TechManNews.
