DeepSeek V4 Pro with Codex and Claude Code: Endpoint Compatibility Test

Can deepseek-v4-pro be used with Codex and Claude Code? The short answer from a 2026-07-05 Crazyrouter test is:

• It works through OpenAI-compatible POST /v1/chat/completions.
• It did not work as a native Codex-style POST /v1/responses model in this test.
• It worked through Claude Messages compatibility at POST /v1/messages, including a basic tool_use -> tool_result loop.

Test base URL:

https://cn.crazyrouter.com/v1

What was tested

GET /v1/models returned HTTP 200 and listed deepseek-v4-pro:

{
  "id": "deepseek-v4-pro",
  "supported_endpoint_types": ["openai"],
  "public_endpoint_types": ["openai"]
}

That matters because "OpenAI-compatible" often means chat completions compatibility, not full Responses API compatibility.

Result table

Test	Endpoint	HTTP	Result
Model list	GET /v1/models	200	deepseek-v4-pro present, endpoint type openai
Chat completion	POST /v1/chat/completions	200	Returned DS_V4_PRO_CHAT_OK with normal token budget
Small token budget	POST /v1/chat/completions	200	Visible output truncated to DS; reasoning tokens used 28/30 completion tokens
Codex-style Responses	POST /v1/responses	400	Failed with convert_request_failed
Claude Messages text	POST /v1/messages	200	Returned Claude-style text content block
Claude Messages tool use	POST /v1/messages	200	Returned tool_use get_city_timezone({"city":"Beijing"})
Claude tool result continuation	POST /v1/messages	200	Accepted tool_result and returned final text

Why Codex is different

Codex is not only a chat client. A custom provider often uses Responses API semantics:

model_provider = "custom"
model = "deepseek-v4-pro"

[model_providers.custom]
wire_api = "responses"
base_url = "https://your-router.example/v1"

That means Codex may expect Responses-style output items, tool calls, tool results, reasoning state, and streaming events. A model that only behaves well behind /v1/chat/completions is not automatically a good Codex backend.

In this test, deepseek-v4-pro was online and usable as a chat model, but POST /v1/responses returned HTTP 400. So the practical conclusion is: do not configure it as a direct Codex Responses model unless your gateway has verified Responses translation.

Why Claude Code can work

Claude Code commonly talks in Anthropic Messages format:

Claude Code -> /v1/messages -> gateway -> /v1/chat/completions -> DeepSeek V4 Pro

The gateway can convert Claude content blocks and tool_use into OpenAI-compatible chat requests, then convert the model output back into Claude-style content blocks.

The test confirmed that this path worked for text output, tool use, and tool-result continuation. That does not make it native Claude support, but it is a practical compatibility path.

Production lesson

Do not judge agent compatibility from model name alone. Test the endpoint contract:

• For Codex: verify /v1/responses, tool call items, tool result continuation, streaming events, and non-empty final output.
• For Claude Code: verify /v1/messages, tool_use, tool_result, SSE/content blocks, usage fields, and retries.
• For DeepSeek V4 Pro: give enough output budget because reasoning tokens can consume part of max_tokens.

Full article: https://crazyrouter.com/blog/deepseek-v4-pro-codex-claude-code-compatibility-2026?utm_source=qiita&utm_medium=article&utm_campaign=deepseek_v4_pro_codex_claude_code_20260705&utm_content=qiita_deepseek-v4-pro-codex-claude-code_20260705__canonical&utm_term=DeepSeek%20V4%20Pro%20Codex%20Claude%20Code