DEV Community

Charles
Charles

Posted on

DeepSeek V4 Pro 0813 Quietly Released — and It Brings OpenAI-Compatible Responses API

#ai #deepseek #machinelearning #opensource

DeepSeek just released V4 Pro 0813, and while it is generating buzz on Hacker News (245 points), the bigger story might be what came with it: a Responses API that is compatible with OpenAI's format, making it trivially easy to swap DeepSeek into applications built for OpenAI.

What Is DeepSeek V4 Pro 0813?

DeepSeek has been on a remarkable trajectory. Earlier this year, DeepSeek V4 Flash made headlines by acing the ARC-AGI-1 reasoning benchmark at $0.02 per task. Now V4 Pro 0813 brings the same architecture to a more powerful configuration.

The model is available through OpenRouter and directly via DeepSeek's API at https://api.deepseek.com. Based on the API documentation, it supports:

  • Standard chat completions (OpenAI-compatible)
  • Responses API (OpenAI's newer format, used by Codex)
  • Streaming with semantic SSE events
  • Tool calls and function calling
  • Thinking mode (extended reasoning)
  • Context caching for reduced costs

The Responses API: Why It Matters

The Responses API is OpenAI's newer interface format that goes beyond simple chat completions. It supports:

  • Instructions — system-level directives separate from user input
  • Streaming events — granular events for reasoning, tool calls, and content parts
  • Built-in tools — web search, code execution, etc.
  • Multi-turn conversations with proper state management

DeepSeek implementing this format means:

  1. Drop-in compatibility with Codex — OpenAI's coding agent can now use DeepSeek as its backend
  2. Easy migration from OpenAI — change the base URL and API key, keep everything else
  3. Access to DeepSeek pricing — which has historically been dramatically cheaper than OpenAI

Code Example

Here is how simple it is to use DeepSeek with the Responses API:

from openai import OpenAI

client = OpenAI(
    api_key="<your DeepSeek API Key>",
    base_url="https://api.deepseek.com"
)

response = client.responses.create(
    model="deepseek-v4-flash",
    instructions="You are a helpful assistant.",
    input="Hi, how are you?",
)

print(response.output_text)
Enter fullscreen mode Exit fullscreen mode

And with streaming:

stream = client.responses.create(
    model="deepseek-v4-flash",
    instructions="You are a helpful assistant.",
    input="Hi, how are you?",
    stream=True,
)

for event in stream:
    if event.type == "response.output_text.delta":
        print(event.delta, end="")
Enter fullscreen mode Exit fullscreen mode

The Streaming Event System

DeepSeek's implementation includes a rich set of streaming events:

  • response.created — response has been created, status in_progress
  • response.in_progress — response is being generated
  • response.output_item.added/done — output item (reasoning/message/function_call) starts/completes
  • response.content_part.added/done — content part within an output item starts/completes
  • response.reasoning_text.delta — reasoning text chunk
  • response.output_text.delta — output text chunk
  • response.completed/incomplete/failed — final event

This granularity is important for building responsive UIs and agent systems where you need to show reasoning in real-time.

The Pricing War

DeepSeek has consistently undercut OpenAI and Anthropic on pricing. With V4 Flash costing roughly $0.02 per task on reasoning benchmarks, the economics of AI are shifting:

  • For developers: You can build production AI applications at a fraction of the cost
  • For startups: The barrier to entry for AI-powered products keeps dropping
  • For hobbyists: Running AI agents (like I do) becomes economically viable even at scale

What This Means for the AI Ecosystem

The release of V4 Pro 0813 alongside the Responses API compatibility tells us several things:

  1. DeepSeek is targeting OpenAI's developer base directly — not just matching capabilities, but matching APIs
  2. The open weights movement is accelerating — DeepSeek's models are open weight, meaning you can self-host if you have the hardware
  3. Competition is driving innovation — with Qwen3.8-2.4T also released today, the pace of releases is unprecedented

Implications for Edge AI

While V4 Pro requires significant hardware to run locally, the architecture and techniques will trickle down to smaller models. The Responses API format is particularly relevant for agent builders — it provides a clean interface for tool use, reasoning, and streaming that works equally well on cloud APIs and local models.

For those of us running AI on edge devices like Raspberry Pi, the key takeaway is that the API format is converging. Building your agent against the Responses API format means you can swap between cloud models (DeepSeek, OpenAI) and local models (Ollama-compatible) with minimal code changes.

DeepSeek V4 Pro 0813 is available via OpenRouter and the DeepSeek API.

Top comments (0)