DeepSeek now supports the Responses API format, with the base URL set to https://api.deepseek.com. To integrate, install the OpenAI SDK and configure the client with your DeepSeek API key. A simple configuration enables using DeepSeek models, such as deepseek-v4-flash or deepseek-v4-pro, in Codex.

Streaming is enabled by setting `stream: true`, returning a sequence of SSE events. The stream ends with events like `response.completed`, `response.incomplete`, or `response.failed`. Key events include `response.output_text.delta` for incremental output and `response.reasoning_text.delta` for chain-of-thought text.

Compatibility details show that many top-level parameters like `model`, `input`, `instructions`, `stream`, `temperature`, `top_p`, and `max_output_tokens` are fully supported. Parameters such as `tools` (function and web_search only), `tool_choice`, and `reasoning` are partially supported. Unsupported parameters like `previous_response_id`, `conversation`, and `store` are silently ignored, allowing existing Responses API clients to connect without modification. Input items support `message`, `function_call`, `function_call_output`, `reasoning`, and `web_search_call` types. The response object mirrors the OpenAI Responses API structure with fixed values for unsupported fields. Token usage is returned with breakdowns for cached and reasoning tokens.