15:07Machine-translated from Azerbaijani by AIRead in Azerbaijani
DeepSeek AI model fails to substantiate claim that it is 80 times cheaper than Claude
“Call Center Doctors” rented a server with 4 “Nvidia H200” accelerators on September 27. The company tested the “DeepSeek V4.1 Flash” model on real-world coding work. The goal was to verify the claim that the model was 80 times cheaper than “Claude.” The server was able to generate approximately 213 tokens per second. A token is a small piece of text processed by an artificial intelligence model. The rental cost $440.88 per day (approximately 750 manats). The same work would cost between $184 and $223 through the “DeepSeek” API. That is approximately 310 to 380 manats.
During off-peak hours, 1 million new input tokens cost $0.15 through the “DeepSeek” API. That is approximately 0.26 manats. A cached, or previously stored, set of 1 million input tokens costs $0.003. The price in manats is 0.005 manats. The 1 million output tokens generated by the model cost $0.60. That is also approximately 1 manat. Prices double during peak hours. For “Claude Opus 5.5,” 1 million input tokens cost $4. That is approximately 7 manats. The same volume of output tokens costs $20. In manats, that is approximately 34 manats. Text read from the cache costs $0.20, or 0.34 manats. The company conducting the test used a cheaper subscription rather than the “Claude” API. According to the test records, 96% of the material the agents sent to the model consisted of earlier text. For every new token, approximately 1,000 old tokens were reread. As a result, most of the server’s time was spent processing the earlier conversation rather than generating new text. It took 5 attempts to get the model running reliably. Each load took 10–15 minutes. Rental charges continued to accrue even when the server was idle. In 1-minute tests during which the server operated at full capacity, it delivered adequate results only when handling 1 type of task at a time. The on-demand rate was $18.37 per hour, or approximately 31 manats. At this price, the 1-month cost would reach approximately $13,200. The amount in manats is 22,440 manats. This cost is more than twice what the company paid for “Claude” in September. The server was able to process approximately 20 billion tokens per day, while the agents processed 51 billion tokens on their busiest day. The data for the company’s main server covers the period from September 1 to September 27. During this period, 2.03 million requests were sent to the model. A total of 388.5 billion tokens were read, 374.2 billion of which came from the cache. The model generated 393 million tokens, and 5,610 changes were merged. The September cost of the “Claude Code” subscriptions was $5,500, or 9,350 manats. The same usage would have cost between $3,500 and $7,000 through the “DeepSeek” API. In manats, this amount ranges from 5,950 to 11,900 manats. The average cost would have been $4,200, or 7,140 manats.
Processing those tokens at the public API prices for the “Claude Opus 5.5” model would have cost approximately $140,000. That is 238,000 manats. The amount is more than 25 times the subscription cost and shows why the claim of being 80 times cheaper did not hold up in real-world work. Each merged change cost approximately $1, or 2 manats, with “Claude.” In the best-case scenario, the cost for “DeepSeek” was calculated at $0.63, or 1 manat. Under a more realistic estimate, the amount ranged from $1.15 to $4.90. In manats, the cost ranged from 2 to 8 manats. The reasons were additional tokens, poor results and the need for rechecking. “DeepSeek” did not write actual code during the test. The checks identified ways the code could escape the closed testing environment. Therefore, 48–64 agents operated in read-only mode. They examined 2,377 directories and prepared 32 bug reports. The server was rented at a discounted rate of $9.19 per hour, or approximately 16 manats. The server was reclaimed a few minutes after the test ended. A release date has not yet been announced for the “DeepSeek V4.1-Pro” model, which is expected to be more powerful.
During off-peak hours, 1 million new input tokens cost $0.15 through the “DeepSeek” API. That is approximately 0.26 manats. A cached, or previously stored, set of 1 million input tokens costs $0.003. The price in manats is 0.005 manats. The 1 million output tokens generated by the model cost $0.60. That is also approximately 1 manat. Prices double during peak hours. For “Claude Opus 5.5,” 1 million input tokens cost $4. That is approximately 7 manats. The same volume of output tokens costs $20. In manats, that is approximately 34 manats. Text read from the cache costs $0.20, or 0.34 manats. The company conducting the test used a cheaper subscription rather than the “Claude” API. According to the test records, 96% of the material the agents sent to the model consisted of earlier text. For every new token, approximately 1,000 old tokens were reread. As a result, most of the server’s time was spent processing the earlier conversation rather than generating new text. It took 5 attempts to get the model running reliably. Each load took 10–15 minutes. Rental charges continued to accrue even when the server was idle. In 1-minute tests during which the server operated at full capacity, it delivered adequate results only when handling 1 type of task at a time. The on-demand rate was $18.37 per hour, or approximately 31 manats. At this price, the 1-month cost would reach approximately $13,200. The amount in manats is 22,440 manats. This cost is more than twice what the company paid for “Claude” in September. The server was able to process approximately 20 billion tokens per day, while the agents processed 51 billion tokens on their busiest day. The data for the company’s main server covers the period from September 1 to September 27. During this period, 2.03 million requests were sent to the model. A total of 388.5 billion tokens were read, 374.2 billion of which came from the cache. The model generated 393 million tokens, and 5,610 changes were merged. The September cost of the “Claude Code” subscriptions was $5,500, or 9,350 manats. The same usage would have cost between $3,500 and $7,000 through the “DeepSeek” API. In manats, this amount ranges from 5,950 to 11,900 manats. The average cost would have been $4,200, or 7,140 manats. Processing those tokens at the public API prices for the “Claude Opus 5.5” model would have cost approximately $140,000. That is 238,000 manats. The amount is more than 25 times the subscription cost and shows why the claim of being 80 times cheaper did not hold up in real-world work. Each merged change cost approximately $1, or 2 manats, with “Claude.” In the best-case scenario, the cost for “DeepSeek” was calculated at $0.63, or 1 manat. Under a more realistic estimate, the amount ranged from $1.15 to $4.90. In manats, the cost ranged from 2 to 8 manats. The reasons were additional tokens, poor results and the need for rechecking. “DeepSeek” did not write actual code during the test. The checks identified ways the code could escape the closed testing environment. Therefore, 48–64 agents operated in read-only mode. They examined 2,377 directories and prepared 32 bug reports. The server was rented at a discounted rate of $9.19 per hour, or approximately 16 manats. The server was reclaimed a few minutes after the test ended. A release date has not yet been announced for the “DeepSeek V4.1-Pro” model, which is expected to be more powerful.