[Solution] How to enable DeepSeek V4 thinking mode in OpenCode using Anthropic API

[LieGeee]

Problem

When using DeepSeek V4 with thinking mode enabled in OpenCode, you may encounter this error:

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

This happens because:

1. DeepSeek V4 returns reasoning_content (thinking process) in addition to content (final answer)
2. In multi-turn conversations with tool calls, reasoning_content must be passed back to the API
3. The @ai-sdk/openai-compatible package doesn't preserve and pass back reasoning_content

Solution

Use DeepSeek's Anthropic-compatible API instead of the OpenAI-compatible one. The @ai-sdk/anthropic package natively supports reasoning mode (Claude extended thinking).

Configuration

Edit ~/.config/opencode/opencode.json:

"deepseekv4": {
  "npm": "@ai-sdk/anthropic",
  "name": "DeepSeek",
  "options": {
    "baseURL": "https://api.deepseek.com/anthropic",
    "apiKey": "your-deepseek-api-key"
  },
  "models": {
    "deepseek-v4-pro": {
      "name": "DeepSeek-V4-Pro",
      "limit": {
        "context": 1048576,
        "output": 262144
      },
      "options": {
        "thinking": {
          "type": "enabled",
          "budgetTokens": 8192
        }
      }
    }
  }
}

Key Changes

1. npm: Change from @ai-sdk/openai-compatible to @ai-sdk/anthropic
2. baseURL: Change from https://api.deepseek.com to https://api.deepseek.com/anthropic
3. thinking.type: Set to "enabled" to activate thinking mode

Why This Works

• DeepSeek provides two API formats: OpenAI-compatible and Anthropic-compatible
• The Anthropic SDK has native support for reasoning mode (used by Claude's extended thinking)
• By switching to the Anthropic API format, OpenCode can properly handle reasoning_content without code changes

Usage

After restarting OpenCode, DeepSeek V4 will work with thinking mode enabled. You'll see the reasoning process in the model's responses.

References

• Community solution article: https://svtter.cn/p/如何解决-opencode-中-deepseek-模型的-reasoning-问题/
• DeepSeek API docs: https://api-docs.deepseek.com/zh-cn/

Tested Environment

• OpenCode version: Latest (as of 2026-04-24)
• DeepSeek API: v4-pro
• OS: Windows 11

[github-actions]

This issue doesn't fully meet our contributing guidelines.

What needs to be fixed:

• This issue doesn't use one of the required templates (Bug Report, Feature Request, or Question). It appears to be a workaround/solution post, which doesn't fit the issue tracker format. If you want to share this as a bug report about missing reasoning_content support, please use the Bug Report template. If you're requesting a fix, use the Feature Request template (with a [FEATURE]: title prefix).

Please edit this issue to address the above within 2 hours, or it will be automatically closed.

---

This issue might be a duplicate of existing issues. Please check:

• Support DeepSeek V4 Thinking Mode with reasoning_content handling #24114: Support DeepSeek V4 Thinking Mode with reasoning_content handling
• DeepSeek thinking mode: reasoning_content must be passed back to API on conversation continuation #24104: DeepSeek thinking mode: reasoning_content must be passed back to API on conversation continuation
• deepseek-v4-pro model ID is not recognized as a thinking model, causing reasoning_content errors even with interleaved configured #24097: deepseek-v4-pro model ID not recognized as thinking model, causing reasoning_content errors
• Missing reasoning_content field in the assistant message for DeepSeek API implementation #17523: Missing reasoning_content field in assistant message for DeepSeek API implementation
• Interleaved reasoning_content lost for non-Claude models on Anthropic SDK #14638: Interleaved reasoning_content lost for non-Claude models on Anthropic SDK

If you believe this was flagged incorrectly, please let a maintainer know.

[AutumnSun1996]

@ai-sdk/anthropic still fails sometimes. here's a better workaround i'm using now:
using transformRequestBody to add all missing strings

      if (body.messages) {
        for (const msg of body.messages) {
          if (msg.role === 'assistant') {
            if (msg.tool_calls && msg.tool_calls.length > 0) {
              if (!msg.content) {
                msg.content = 'call tool';
              }
              if (!msg.reasoning_content) {
                msg.reasoning_content = 'call tool';
              }
            } else {
              if (!msg.content) {
                msg.content = 'wait user input';
              }
              if (!msg.reasoning_content) {
                msg.reasoning_content = 'answer';
              }
            }
          }
        }
      }

[LieGeee]

使用和token记录

[LieGeee]

@ai-sdk/anthropic still fails sometimes. here's a better workaround i'm using now:@ai-SDK/Anthropic 有时还是会失败。这里有我现在用的更好的变通方法： using transformRequestBody to add all missing strings使用 transformRequestBody 添加所有缺失字符串

  if (body.messages) {
    for (const msg of body.messages) {
      if (msg.role === 'assistant') {
        if (msg.tool_calls && msg.tool_calls.length > 0) {
          if (!msg.content) {
            msg.content = 'call tool';
          }
          if (!msg.reasoning_content) {
            msg.reasoning_content = 'call tool';
          }
        } else {
          if (!msg.content) {
            msg.content = 'wait user input';
          }
          if (!msg.reasoning_content) {
            msg.reasoning_content = 'answer';
          }
        }
      }
    }
  }

@ai-sdk/anthropic still fails sometimes. here's a better workaround i'm using now:@ai-SDK/Anthropic 有时还是会失败。这里有我现在用的更好的变通方法： using transformRequestBody to add all missing strings使用 transformRequestBody 添加所有缺失字符串

  if (body.messages) {
    for (const msg of body.messages) {
      if (msg.role === 'assistant') {
        if (msg.tool_calls && msg.tool_calls.length > 0) {
          if (!msg.content) {
            msg.content = 'call tool';
          }
          if (!msg.reasoning_content) {
            msg.reasoning_content = 'call tool';
          }
        } else {
          if (!msg.content) {
            msg.content = 'wait user input';
          }
          if (!msg.reasoning_content) {
            msg.reasoning_content = 'answer';
          }
        }
      }
    }
  }

是的这个方案更佳

[LieGeee]

Closing to recreate with proper template format.

[CycSpring]

在openclaw调用deepseekV4的时候，我也遇到了；openclaw.json里应该怎么改？

[LieGeee]

上面有

[arjunp1120]

@ai-sdk/anthropic still fails sometimes. here's a better workaround i'm using now: using transformRequestBody to add all missing strings

  if (body.messages) {
    for (const msg of body.messages) {
      if (msg.role === 'assistant') {
        if (msg.tool_calls && msg.tool_calls.length > 0) {
          if (!msg.content) {
            msg.content = 'call tool';
          }
          if (!msg.reasoning_content) {
            msg.reasoning_content = 'call tool';
          }
        } else {
          if (!msg.content) {
            msg.content = 'wait user input';
          }
          if (!msg.reasoning_content) {
            msg.reasoning_content = 'answer';
          }
        }
      }
    }
  }

What does that mean?? where do you use that? i defined my plugin but it doesn't work; can you please clarify what you mean by this code???

[LieGeee]

你发给ai吧,我已经开始用pi了,很久没用opencode了不记得,你有web_search吗,直接把链接发过去,没有就直接粘贴吧,应该ai可以改的

[CycSpring]

哈哈，挺久前的问题了；AI几天一个样，我现在都用的codex、cc、hermes了