- The author says they used DeepSeek 4.1 Flash heavily for about a month across several projects. Without looking at the model name during a chat, they found it hard to tell apart from Opus. This is a subjective impression, and benchmarks are only mentioned as a link in the original.
- The author says a monthly OpenCode Go subscription of 10 makes usage practically close to unlimited. The cost of a file organization task was about 0.003, and most sessions barely exceeded 1.
- For important work, they get a final code review from Opus 5.5 and leave fixes to DeepSeek. The author explains that the reason to call the higher-tier model is to get a different perspective rather than better quality.
- The author claims DeepSeek reduced the KV cache by about 437 times compared with the previous V1 model. The KV cache is data kept in GPU memory while processing conversation content, and the author says it determines the cost of long sessions. The source of this figure and its measurement conditions are not given in the original.
Rather than the feeling that it was nearly indistinguishable from Opus for a month, what stands out is the math: at 10 a month, even using it all day rarely goes past 1. But the source does not say under what conditions the 437-fold cache reduction was measured, so I want to separate simple tasks like file cleanup from long coding sessions and check the actual costs myself.