DeepSeek V4 in OpenCode: Setup Guide

DeepSeek officially documents OpenCode as a supported integration for V4 models. Setup is one command — but a thinking-mode gotcha will break tool-heavy sessions on the default OpenAI-compatible provider unless you fix it first. This guide covers both.

Why OpenCode

OpenCode is an open-source, terminal-native coding agent that runs in your repo, reads AGENTS.md, and drives multi-turn tool use the same way Claude Code does — edit files, run commands, review diffs — all locally, with the model of your choice behind it. Because it's model-agnostic, it's a natural home for DeepSeek V4, and DeepSeek's documentation lists it as an officially supported integration alongside Claude Code and others.

Quick Setup

Use OpenCode 1.14.24 or newer — DeepSeek explicitly recommends the latest version to avoid compatibility issues.

1. Install OpenCode
2. Run opencode, type /connect, enter deepseek
3. Paste your DeepSeek API key
4. Select DeepSeek-V4-Pro

Model identifiers:

OpenCode's model picker also surfaces DeepSeek models through its built-in Model Hub; the identifiers above are the ones to use in config. V4 Pro supports 1M-token context in its [1m] variant (the standard model ships with a 262,144-token context window) — relevant if your agent loop needs to hold a large repository or long session history without churning cache.

The Thinking-Mode Bug

DeepSeek's thinking mode returns a reasoning_content field alongside content — the model's chain of thought. In multi-turn tool calls, that field must be passed back to the API unchanged. When it isn't, DeepSeek rejects the request with:

400: The reasoning_content in the thinking mode must be passed back to the API

OpenCode's default DeepSeek provider uses @ai-sdk/openai-compatible, which does not reliably preserve and replay reasoning_content across turns. Agentic sessions are nothing but turns, so this surfaces on the second or third tool call. The same class of bug affects V4 Flash.

Fix 1: Route Through the Anthropic-Compatible Endpoint (Recommended)

The Anthropic SDK natively understands reasoning blocks — Claude extended thinking works the same way — so OpenCode stops dropping the field. Add a provider block in ~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "deepseek-anthropic": {
      "npm": "@ai-sdk/anthropic",
      "name": "DeepSeek (Anthropic-compatible)",
      "options": {
        "baseURL": "https://api.deepseek.com/anthropic",
        "apiKey": "{env:DEEPSEEK_API_KEY}"
      },
      "models": {
        "deepseek-v4-pro[1m]": {
          "name": "DeepSeek V4 Pro (1M)",
          "limit": { "context": 1048576, "output": 262144 },
          "options": { "thinking": { "type": "enabled", "budgetTokens": 8192 } }
        },
        "deepseek-v4-flash": {
          "name": "DeepSeek V4 Flash"
        }
      }
    }
  }
}

Restart OpenCode and thinking mode works across turns. The budgetTokens line caps the reasoning budget; raise it for hard problems, lower it for routine work to save tokens and latency. If you prefer the native OpenAI-compatible API but still want thinking, the @ai-sdk/openai-compatible route requires passing reasoning_content back in every turn — the Anthropic endpoint removes that footgun entirely.

Fix 2: Disable Thinking on OpenAI-Compatible Endpoints

If you want to stay on DeepSeek's native OpenAI-compatible API, disable thinking for the model:

{
  "models": {
    "deepseek-v4-pro": {
      "options": {
        "extra_body": { "thinking": { "type": "disabled" } }
      }
    }
  }
}

This is stable on plain OpenAI-compatible providers, but the cost is explicit: you lose the visible reasoning that makes V4 Pro strong on hard problems. If you're disabling thinking for everyday work, V4 Flash deserves a hard look — it's the cheaper, faster model for routine agent tasks.

Pricing Note

DeepSeek moved from flat pricing to peak/off-peak rates on August 16, 2026. V4-Pro: $0.66 in / $1.98 out per 1M tokens off-peak; $1.32 in / $3.96 out at peak. Cache-hit input is dramatically cheaper ($0.022–$0.044). Peak hours are 01:00–04:00 and 06:00–10:00 UTC. Model a budget around cache hits and off-peak scheduling — for agent loops that repeatedly retrieve the same context, cache stability is the deciding cost factor. The same economics apply through the Anthropic-compatible endpoint, so Fix 1 doesn't change your cost model — it changes your reliability.

Performance Expectations

Independent runs show V4 Pro is harness-sensitive: it scored 98+ under DeepSeek's own setup but 91–96 under OpenCode's default system prompt (modeltest, Aug 14). If OpenCode with V4 Pro feels a notch below the frontier, the harness is the likely culprit, not the model. Two things to try before switching models: use a reasoning-heavy system prompt tuned for the task, and consider the Code-style agent pattern — collapsing multi-round-trip sequences into single programmatic calls reduces the number of turns where harness overhead can bite. See the DeepSeek Harness blog for the full picture.

Troubleshooting

• "Reasoning_content must be passed back" error — you're on the OpenAI-compatible provider with thinking enabled. Apply Fix 1 (Anthropic-compatible endpoint), or disable thinking with Fix 2.
• Model not listed in the picker — after /connect deepseek, verify the model IDs in your config match the provider's catalog (deepseek-v4-pro, deepseek-v4-pro[1m], deepseek-v4-flash). Custom model blocks you've added with a models key override the picker list — check opencode.json.
• 401 unauthorized on first turn — the API key env var isn't reaching the process. Confirm DEEPSEEK_API_KEY is exported in the shell that launches opencode, not just written to a file.
• V4 Flash feels wrong for the task — Flash is the fast/cheap tier; hard problems are V4 Pro's job. With thinking disabled, use Flash for routine agent turns and switch to Pro (thinking on) for anything requiring visible reasoning.
• After a config change, nothing updates — restart OpenCode. Provider blocks and model options are read at startup.

OpenCode stores session state locally, and AGENTS.md files in your repo are picked up automatically — so rules tuned with the agentic prompt maintenance guidance apply to V4 sessions the same way they do to any other model.

Related

• Claude Code Integration — run DeepSeek V4 inside Claude Code
• OpenAI-Compatible API — native API reference
• Tool Calls with Thinking — handling reasoning_content with tool calls
• DeepSeek Harness — DeepSeek's own open-source agent runtime