API
< Back to all clusters
[TECHNOLOGY] · China · 38 sources

started · updated

DeepSeek launches V4 Pro flagship AI model with enhanced agent capabilities

DeepSeek has officially released the general availability version of its flagship AI model, DeepSeek-V4-Pro (build 0813). The model utilizes a Mixture-of-Experts (MoE) architecture with approximately 1.6 trillion total parameters, activating about 49 billion per request. Key technical features include a 1-million-token context window, support for up to 384,000 output tokens, and native compatibility with OpenAI’s Responses API format to facilitate integration with agentic tools like Codex.

In benchmark testing, the V4 Pro model demonstrated significant improvements in agentic capabilities, scoring 87.9 on the Terminal Bench 2.1, closely approaching Anthropic’s Claude Fable 5. The company is also expanding its ecosystem with the DeepSeek Harness framework, an open-source tool designed to enhance AI agent workflows.

Alongside the launch, DeepSeek is implementing a major shift in its pricing strategy. Starting August 16, the company will introduce a peak and off-peak pricing model for its APIs to manage computational resources more effectively. During peak hours, V4 Pro output costs will rise significantly, while off-peak rates will be half the peak price. Despite these increases, the company’s models remain highly competitive in cost compared to Western counterparts like OpenAI and Anthropic.

Entities

Alibaba · Anthropic · Beijing · ByteDance · Claude · DeepSeek · DeepSeek Harness · Hangzhou · OpenAI

Claims

What the coverage asserts, and how many sources carry each claim.

Sources

DeepSeek宣布涨价,大模型的竞争方式将被改写?_天极网 [yesky.com]
一个入口用多个国内大模型:GLM、DeepSeek、LongCat 分别擅长什么_天极网 [news.yesky.com]