DeepSeek rolls out V4 AI model, claims edge over key U.S. rivals

Chinese AI startup DeepSeek on Friday unveiled preview versions of its V4 model, claiming improvements in reasoning and agentic capabilities that rival or surpass several leading American AI products.

The release intensifies the China-U.S. AI rivalry, with DeepSeek's benchmarks aimed directly at products from OpenAI, Google and Anthropic — tools many Korean-American technology workers and entrepreneurs in Orange County rely on daily.

DeepSeek's V4 comes in "pro" and "flash" versions, both featuring 1 million token context windows — a substantial expansion from V3's 128,000-token limit. The company released V3 in late 2024 and its R1 reasoning model in January 2025. Both V4 versions are described as open-source, allowing developers to modify and build on the underlying technology.

The company claims V4 Pro Max outperforms OpenAI's GPT-5.2 and Google's Gemini 3.0-Pro but falls marginally short of GPT-5.4 and Gemini 3.1-Pro, according to the Korea Times. DeepSeek also said V4 Pro's agentic capabilities could outperform Anthropic's Claude Sonnet 4.5 and approach Claude Opus 4.5.

"Based on the benchmark results, it does appear DeepSeek V4 is going to be very competitive against its U.S. rivals," said Lian Jye Su, chief analyst at Omdia.

Marina Zhang, an associate professor at the University of Technology Sydney, described the rollout as a "pivotal milestone for China's AI industry." The V4 preview versions are supported by Huawei's Ascend chips, reducing DeepSeek's dependence on Nvidia processors. "Domestic competition has intensified significantly since R1's release," said Ivan Su, a senior equity analyst at Morningstar.

DeepSeek's rapid advances have drawn scrutiny from American rivals. Anthropic publicly accused the company in February 2026 of running industrial-scale campaigns to extract AI capabilities — a practice the company described as one that "involves training a less capable model on the outputs of a stronger one," according to the Korea Times. OpenAI leveled similar allegations in a letter to U.S. lawmakers. China's embassy in Washington rejected the claims as "unjustified suppression of Chinese companies by the U.S."

DeepSeek's V4 follows its R1 reasoning model, released in January 2025, which drew global attention for matching U.S. competitors at a fraction of the reported development cost.