A terminal coding agent that costs nothing to try sounds like an easy decision. The catch sits in the details: which model is actually free, for how long, what happens to your code, and whether the setup you copy matches the current documentation. OpenCode with DeepSeek V4.1 Flash Free is a good example of why those details matter.
What OpenCode is, and what it is not
A coding agent is a program that takes a request in plain language, reads your project files, writes or edits code and can run commands to check the result. OpenCode describes itself as an open source coding agent that runs as a terminal interface, a desktop app or a web app. The terminal is the text window where you type commands instead of clicking.
The working style is close to Claude Code: you describe a task, the agent plans, edits files and reports back. But OpenCode is a separate project from a different team. It is not Claude Code under another name, and it does not run Anthropic’s models unless you connect them yourself. What it borrows is the idea and, as we will see, part of the folder structure.
Is DeepSeek V4.1 Flash really free in OpenCode?
Yes, but only one specific entry. OpenCode’s model list contains two DeepSeek options that look almost identical. “DeepSeek V4.1 Flash” is a paid model at $0.30 per million input tokens and $1.20 per million output tokens. A token is a small piece of text, roughly part of a word. “DeepSeek V4.1 Flash Free”, with the ID deepseek-v4.1-flash-free, is listed as free for input, output and cached reads.
The documentation adds three conditions. The free entry is available “for a limited time”. Its provider follows a zero-retention policy and does not use your data for model training. And free-tier chat models only work inside the OpenCode client, not through a direct API or another tool. The page gives no numeric usage limits.
Small detail. Big difference. If you pick the entry without “Free” in its name, you are on the metered model. And the free list is not uniform: some other free models in the same list state that collected data may be used to improve the model. If your project contains client code or anything confidential, that line is worth reading before the first prompt. Teams that prefer not to see free models at all can hide them with the environment variable OPENCODE_DISABLE_FREE_MODELS=1.
What DeepSeek V4.1 Flash is
DeepSeek released V4.1-Flash on 10 September 2026. Its API name is deepseek-flash. DeepSeek calls it the smallest model in its new architecture family, with native visual understanding. It is a mixture-of-experts model, meaning only part of the network works on each request: 552 billion parameters in total, with 8 billion active for input and 16 billion for output. DeepSeek also says the model needs a quarter of the memory for its working cache compared with the previous generation.
DeepSeek claims benchmark results ahead of its own former flagship V4-Pro, and it now routes V4-Pro requests to V4.1-Flash. Those are vendor numbers. They tell you where to look, not how the model will behave on your codebase.
How to install OpenCode and choose the free model
The official documentation lists several install methods. The script install is curl -fsSL https://opencode.ai/v2/install | bash. If you already use Node.js, npm install -g @opencode/cli works too, and Homebrew, Bun, pnpm and other package managers are listed.
A sensible order looks like this. First, create a folder for the project and start OpenCode inside it, so the agent only sees files that belong to the task. Second, connect a provider with /connect in the interface. The documentation describes signing in to the OpenCode Console and pasting an API key; it does not say clearly whether the free models skip that step, so plan for a sign-in. Third, run /models and select DeepSeek V4.1 Flash Free. In a config file the same model is written as opencode/deepseek-v4.1-flash-free.
Can you reuse your Claude Code skills?
Largely, yes. A skill is a Markdown instruction file the agent loads when a task matches it. OpenCode searches its own folders, ~/.config/opencode/skills and .opencode/skills, and also the Claude-compatible paths ~/.claude/skills and .claude/skills, plus .agents/skills. A nested skill must be a file named exactly SKILL.md in its own directory. The description in the frontmatter, the short header block at the top of the file, is what the model sees; a skill without a description is not offered to the model.
That makes switching cheaper, but it does not make the result identical. The same instructions are interpreted by a different model, so a skill that works well with one model deserves a fresh test with the other.
How to build your own helper agent
OpenCode lets you define custom agents as Markdown files. For one project they live in .opencode/agents/<name>.md; for all projects in ~/.config/opencode/agents/. The body of the file becomes the agent’s system prompt, its standing instructions. In the frontmatter, description explains the purpose and mode sets the role: primary runs as the main agent of a session, subagent runs in a child session when the main agent calls it, and all allows both. A permissions list can allow, ask for or deny specific actions.
A typical beginner use is an error explainer: an agent that reads the project, lists each fault in plain words and proposes a fix. Before the agent fixes anything, write down what the page or tool should do. Then check the result against that description. Code that the agent wrote and code that does what you intended are not the same thing, and only the second one counts.
Where this setup makes sense
For learning, prototypes and small internal tools, a free model inside an open source agent is a low-risk way to find out whether agent-based coding suits you. Two points argue for caution. “Limited time” means the free entry can disappear, so do not build a routine that depends on it. And folder names change between versions: the documentation currently uses the plural agents and skills directories, so older tutorials with other paths may silently do nothing. When a setting seems to have no effect, compare your path with the docs before you blame the model.
