DeepSeek V4 Pro 0813 adds OpenAI-compatible Responses API with streaming and tool support

DeepSeek V4 Pro 0813 introduces an OpenAI-compatible Responses API that allows existing OpenAI SDK clients to switch endpoints with a single base_url change. The API accepts standard OpenAI parameters like model, input, and stream, but silently drops unsupported ones such as conversation history or metadata. This design choice simplifies integration for developers already using OpenAI’s ecosystem but removes features that rely on stateful interactions or custom context management.

The streaming implementation provides fine-grained event tracking for output generation, including incremental text deltas, reasoning steps, and tool call progress. Events like response.output_text.delta or response.function_call_arguments.delta enable real-time processing of model outputs, while final events (completed/incomplete/failed) include usage data or error details. However, the API lacks support for parallel tool calls and enforces a stateless model, which may limit applications requiring multi-turn interactions or complex tool orchestration.

Tool support is partial but functional, with function calls and web search available while other tool types are ignored. The API accepts tool_choice parameters but does not enforce max_tool_calls, defaulting to parallel execution. Web search events provide status updates (in_progress/searching/completed), but the absence of conversation state or context caching means developers must manage these aspects client-side. The stateless nature also prevents features like prompt caching or service tier selection, which could impact cost or performance optimization strategies.