replaced my $200/mo chatgpt with a $10/mo setup. opencode go plus deepseek harness. deepseek harness is deepseek's own official agent framework, built on a fully pluggable architecture, everything from the llm adapter to the shell sandbox is a swappable plugin, with a real observability dashboard tracking every request down to the token and millisecond. opencode go is a subscription llm gateway that slots straight in, their api conventions match deepseek harness's expectations exactly, so switching is two environment variables, no code changes, no adapter to write. the real economics here. agent loops resend the full conversation history on every step, so input tokens balloon fast, one documented 34-step task hit 1.9 million cumulative prompt tokens against under 10k of actual output. prefix caching hits 96.5% in practice, and a flat subscription turns "can i afford to let the agent run a few more steps" from a per-call cost question into a fixed budget question. worth knowing the honest limits too. deepseek's models are text only, no image input works in the harness yet. no completion notification, you're checking a browser tab manually on long tasks.