DeepSeek于8月13日晚间正式发布V4 Pro 0813版本API[2]。新版本支持Responses API和Codex接入[2]，并增强了Agent能力[2]。该模型在多项测试中的性能表现接近Fable 5水平[2]。
新API基础URL为https://api.deepseek.com[1]，支持deepseek-v4-flash模型[1]，兼容OpenAI SDK[1]。该接口支持流式传输[1]，并返回input_tokens、output_tokens、cached_tokens和reasoning_tokens等token用量统计数据[1]。
定价方面，百万tokens输入（缓存命中）为0.025元，百万tokens输入（缓存未命中）为3元，百万tokens输出为6元[2]。
DeepSeek has quietly rolled out version V4 Pro 0813, introducing support for the Responses API format that enables developers to integrate DeepSeek models into Codex through straightforward configuration [1]. The official API launch occurred on the evening of August 13 [2].
The new model version, DeepSeek-V4-Pro-0813, enhances agent capabilities and maintains compatibility with OpenAI's SDK standards [1][2]. The API supports streaming responses and provides detailed token usage statistics, including separate tracking for input tokens, output tokens, cached tokens, and reasoning tokens [1]. Performance benchmarks across multiple tests show the model approaching Fable 5 levels [2].
Pricing for the API has been announced as follows: cached input tokens are charged at 0.025 yuan per million tokens when cache hits occur, uncached input tokens at 3 yuan per million tokens, and output tokens at 6 yuan per million tokens [2]. The base API endpoint is accessible at https://api.deepseek.com, with the deepseek-v4-flash model available for integration [1].
