DeepSeek V4 Pro with OpenCode: Setup & Cost

OpenCode is an open-source terminal coding agent that speaks the OpenAI-compatible protocol — the same contract DeepSeek exposes at https://api.deepseek.com^[1]. DeepSeek's GA notes explicitly list OpenCode as a supported backend client^[2], and V4 Pro 0813 brings near-frontier patch coding (96.40% SWE-bench Verified)^[5].

Because both sides are open source, the pair has no vendor lock-in: swap the model string and the same OpenCode config talks to any OpenAI-compatible provider.

There is a practical reason this pairing gets so much attention in 2026: it is the cheapest way to run a near-frontier coding agent without renting a closed-model quota. OpenCode brings the agent loop (plan, edit, test, repeat) and V4 Pro brings the reasoning — and both are MIT-licensed, so CI pipelines, containers, and air-gapped setups can run the same stack the developer runs locally^[2][6].

OpenCode reads providers from its config file. Add a deepseek provider block pointing at the official base_url, then select the model^[1][2].

• baseURL https://api.deepseek.com — OpenAI-format endpoint^[1].
• apiKey from the DEEPSEEK_API_KEY environment variable.
• 1M context (1048576 tokens) matches the model's real window^[6].

# opencode config (opencode.json / config.json)
{
  "provider": {
    "deepseek": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "DeepSeek",
      "options": {
        "baseURL": "https://api.deepseek.com",
        "apiKey": "{env:DEEPSEEK_API_KEY}"
      },
      "models": {
        "deepseek-v4-pro": {
          "name": "DeepSeek V4 Pro (0813)",
          "limit": { "context": 1048576, "output": 393216 }
        },
        "deepseek-v4-flash": {
          "name": "DeepSeek V4 Flash (0731)",
          "limit": { "context": 1048576, "output": 393216 }
        }
      }
    }
  },
  "model": "deepseek/deepseek-v4-pro"
}

Use deepseek-v4-pro for agent sessions and deepseek-v4-flash for quick edits. DeepSeek's reasoning_effort (low/high/max) is passed through the OpenAI-compatible API, and thinking mode is enabled by default^[3].

OpenCode's agent loop (plan → edit → test → repeat) maps cleanly onto V4 Pro's tool-calling support. Keep the per-session context lean and stable so DeepSeek's automatic prefix caching trims the bill^[4].

# inside OpenCode sessions you can switch per task:
#  /model deepseek/deepseek-v4-pro
#  /model deepseek/deepseek-v4-flash

# reasoning_effort rides in the request body:
#   {"reasoning_effort": "high"}   <- default, daily agent work
#   {"reasoning_effort": "max"}    <- hardest tasks[3]

OpenCode + V4 Pro is one of the cheapest frontier-ish agent stacks: $0.022 per SWE-bench test on V4 Pro 0813 (Vals AI neutral harness)^[5], and API rates of $0.435/$0.87 per 1M before the 8/16 peak/off-peak change (off-peak $0.66/$1.98 after)^[4].

The open-source stack trades about half a point of SWE accuracy for a 58x cost reduction versus the closed flagship^[5]. For teams running thousands of agent iterations a day, that difference is the difference between a $220 bill and a $13,000 bill.

Cache-aware session design matters more here than on closed models: keep the repository context and system prompt byte-identical across iterations so DeepSeek's automatic prefix cache serves the reused prefix at $0.003625 per 1M (pre-8/16) or $0.022 off-peak (post-8/16) instead of the miss price^[4]. OpenCode's built-in session persistence helps — avoid re-sending large file contexts every turn.

DeepSeek also ships its own agent runtime, DeepSeek Harness (dsh), where official benchmarks run in 'minimal mode'^[7]. OpenCode is the lighter, drop-in alternative: no new runtime, just a config block, and it supports any OpenAI-compatible provider — not just DeepSeek.

Start with OpenCode for a zero-friction path to V4 Pro today, and evaluate dsh when its plugin ecosystem (now at 60+ community plugins) stabilizes past the developer-preview stage^[7].