DeepSeek v4 Lands, and the Frontier Just Got Crowded Again

Remember January 2025? DeepSeek R1 dropped, Nvidia lost roughly $600 billion in market cap in a single session, and every AI thread on X spent a week arguing about training costs. Fifteen months later, DeepSeek is back with v4 — and Bloomberg’s April 24 headline (“DeepSeek Unveils AI Model To Challenge Anthropic, OpenAI”) suggests the industry isn’t shrugging this one off either.

The budget killer grows up

DeepSeek made its name on a number: $6 million, the figure they claimed it cost to train an o1-class model with R1. Skeptics picked that apart for months, but the trajectory since has been hard to dismiss. V3.2 reportedly hit 96% of ChatGPT’s score on common benchmarks. R2 quietly extended the lineup. Now v4.

The pitch this time is different. v4 isn’t being sold as another raw LLM win — it’s positioned around agentic capability, the same battleground where Anthropic’s Claude and OpenAI’s GPT-5.5 have been racking up wins. Leaks in March pegged it as a roughly 1-trillion-parameter model, and the launch specs land in that neighborhood.

Open weights are the actual weapon

Strip away the benchmark theater and one thing makes v4 genuinely uncomfortable for the US labs: it’s open source. Claude and GPT-5.5 are API-only. DeepSeek has been shipping weights since R1, and v4 continues that. For an enterprise weighing data residency, EU AI Act compliance, or just the cost of API calls at scale, “frontier-class performance you can run on your own GPUs” is a category that barely existed two years ago.

There’s also a story-arc detail worth flagging. In February, reporting surfaced that DeepSeek had refused Nvidia’s model access — a remarkable inversion of the old narrative where Chinese AI was supposed to be hobbled by export controls. Washington’s chip restrictions were meant to slow China down. Instead they pushed DeepSeek’s team into squeezing more out of less, and the training-efficiency gap has narrowed faster than most US analysts predicted.

Read the benchmarks with sunglasses on

A word of caution. YouTube is already flooded with “DeepSeek v4 Killed GPT-5” thumbnails, most of them recycled from leak-stage content shot before the official release. The January “DeepSeek V4 Leaked” video pulled around 40K views and spawned dozens of imitators. Independent benchmark runs — the kind you’d want before betting a roadmap on this — are barely starting.

And v4 isn’t landing in a vacuum. Baidu’s Ernie, Alibaba’s Qwen, and Tencent’s Hunyuan are all shipping competitive models on overlapping timelines. The “who’s really #1” question is becoming less interesting than the structural one: there are now four or five Chinese labs capable of frontier-adjacent releases, and they’re iterating on monthly cycles.

What this actually changes for builders

If you ship product on top of LLMs, v4 widens your menu in a concrete way. The two-option world — pay OpenAI’s bill or train your own — has been quietly replaced by a third option: take a frontier-grade open weight, host it yourself, control your data. That’s a real architectural choice now, not a hobbyist exercise.

The caveats haven’t gone anywhere. Compliance reviews around Chinese-origin models are still painful in regulated industries. Topic-level censorship and value alignment skew remain documented issues. Technical capability and deployability are not the same conversation.

But the bigger point is this: the question R1 forced last year was “can China catch up?” The question v4 forces today is “can the US stay ahead?” Look at your AI stack. Now picture it six months from now. Not so obvious anymore, is it.