Story perspectives

DeepSeek's v4 Model Debuts, Still Months Behind U.S. AI

4/29/2026

Summary

DeepSeek Launches v4

DeepSeek launched its v4 model on April 24, optimized for Huawei Ascend chips.
Artificial Analysis ranked v4 Pro just behind U.S. models GPT-5.5 and Claude Opus.
Bloomberg reports DeepSeek lags three to six months behind top U.S. systems such as GPT-5.4.
Council on Foreign Relations senior fellow Chris McGuire says U.S. models lead by about seven months.

Summary

DeepSeek Raises Capital

Liang Wenfeng increased his equity stake in DeepSeek to 34 percent, raising his subscribed capital from 100,000 yuan to 5.1 million yuan.
DeepSeek raised its registered capital from 10 million yuan to 15 million yuan as it began external fundraising.
The research and engineering team grew to 270 employees, up from 212 in early December.
Senior researcher Chen Deli posted on X after the V4 launch, stating the team stays true to long-termism and open source for all.

Summary

DeepSeek Faces Competition

DeepSeek’s April 24 launch of model v4 elicited a muted market reaction described as a “shrug.”
Analysts say the lab now faces heightened competition from domestic AI startups and established Western firms.
Growing interference from Chinese state agencies adds regulatory pressure to DeepSeek’s strategy.
Earlier excitement briefly knocked down Nvidia’s market value as DeepSeek’s low-cost models seemed to threaten hardware demand.

Summary

DeepSeek Cuts

DeepSeek cut V4-Pro input price to US$0.0036 per million tokens, about 97 % cheaper than OpenAI.
API “input cache hit” fee fell to one-tenth, lowering the minimum charge to roughly US$0.14 per million tokens.
DeepSeek offers developers a 75 % discount on V4-Pro until May 5, sparking price-war concerns.
CTO Roy Benesh of eSIMple says the cuts let small firms and researchers use AI affordably, likely spurring new ideas.

History

Historical Context

In May 1997 IBM’s Deep Blue beat world champion Garry Kasparov 3½–2½, the first computer to defeat a reigning champion under tournament conditions. That landmark win launched a wave of AI competition that now resurfaces as DeepSeek’s V4 launch, whose 75 % discount fuels a price war with rivals like Nvidia and OpenAI.

1 of 5

DeepSeek Launches v4

DeepSeek launched its v4 model on April 24, optimized for Huawei Ascend chips.
Artificial Analysis ranked v4 Pro just behind U.S. models GPT-5.5 and Claude Opus.
Bloomberg reports DeepSeek lags three to six months behind top U.S. systems such as GPT-5.4.
Council on Foreign Relations senior fellow Chris McGuire says U.S. models lead by about seven months.

1 / 5