The free DeepSeek model in OpenCode is the free version of DeepSeek V4.1 Flash, with the model ID opencode/deepseek-v4.1-flash-free. OpenCode announced on 15 October 2026 that it is free to use inside OpenCode, and the official Zen documentation calls it free for a limited time, with no end date published. Picking the right model is the first step — pick the paid version and you start paying.
Use it in three steps
First, install and open OpenCode, go to the model picker, and find the DeepSeek V4.1 Flash Free entry; you can also run opencode run on the command line with the -m flag to name this model ID. Second, ask questions or hand over code-change tasks directly inside OpenCode — you do not need a separate DeepSeek API key or a card on file, because the free version only responds inside the OpenCode app. Third, when context runs short, mention specific files with and start a fresh session per task instead of stuffing the whole repository into one session.
Three hard limits
The free version's context limit is 200K tokens against 1M for the paid version, and its output limit is 32K, so large-repo tasks hit the ceiling sooner. The free version only works inside OpenCode; calling it as a general-purpose API returns a 403 error saying the free tier can only be used within OpenCode. Speed is not guaranteed either — community reports show generation speed dropping noticeably at large context sizes. DeepSeek's official API itself has no free tier; the paid version in OpenCode Zen is billed at $0.30 per million input tokens and $1.20 per million output tokens, so do not mix up the two routes' prices.
When to switch to the paid version
For everyday code changes and small fixes where 200K of context is enough, the free version is fine to start with. When context keeps filling up, or the model needs to be connected to other tools, switch to the paid version or use DeepSeek's official API directly. In the same period, OpenCode also offers a free GLM 5.3 Flash; when the free DeepSeek version is throttled or queued, you can alternate — its context limit is also 200K.
