DeepSeek V4.1 Flash 1017: OpenCode vs DSH vs direct API speed and cache
A user compares DeepSeek V4.1 Flash 1017 across OpenCode's default route (50–130 tok/s), the direct API (200–250 tok/s) and DeepSeek Harness (200–300 tok/s), with ~99.7–99.8% cache hit rates on Pi + DeepSeek.
i’ve been comparing DeepSeek V4.1 Flash 1017 across OpenCode Go, DSH and the direct API. 5 things stood out: 1. on the default OpenCode route, i’ve seen it drop to around 50–130 tok/s, especially during peak hours. 2. direct DeepSeek API feels much better for me. i’m getting roughly 200–250 tok/s, and tasks often finish with less wasted usage. 3. the harness matters more than i expected. with DeepSeek Harness, i can get the same Go subscription much closer to the 200–300 tok/s range instead of the 50–130 tok/s i was seeing on the default route. 4. cache efficiency is one of the biggest differences i’ve noticed. with Pi + DeepSeek, i can get cache hit rates around 99.7–99.8%, which makes long sessions noticeably cheaper. 5. i’ve stopped looking at “$80 of usage for $10” as the main metric. what matters more to me now is completed tasks per dollar. the same model can feel slow, wasteful or insanely good depending on the harness.