DeepSeek V4

DeepSeek just released its latest flagship model V4, which is massive, powerful, open-source, and a fraction of the cost of competitors.

DeepSeek V4 might be the model that ends America's lead in artificial intelligence, not because China caught up, but because of what happens next.

America has the best chips and the most money flowing into AI labs, yet China was able to release a frontier-level, open-source, open-weights model that matches the best, using nerfed Nvidia GPUs and at a fraction of the cost.

About 18 months ago, DeepSeek released DeepSeek R1, an open-source, open-weights thinking model that showed non-US labs could reach the frontier and caused the stock market to drop 20% overnight.

DeepSeek R1's training efficiency was a fraction of the cost of hundreds of billions spent by US frontier labs, and the resulting belief that Nvidia GPUs were overvalued was wrong because lower prices drive higher demand (Jevons Paradox).

DeepSeek wrote an incredibly thorough white paper for V4, being much more honest about their failures than any closed-source US AI lab.

DeepSeek V4 comes in two flavors, Pro and Flash, and supports a 1-million-token context length, putting it at the frontier of context limits.

DeepSeek V4 Pro is a 1.66-trillion-parameter Mixture of Experts model with 49 billion active parameters; V4 Flash has 284 billion total parameters with 13 billion active; both were trained on about 33 trillion tokens.

DeepSeek V4 has enhanced agentic capabilities comparable to state-of-the-art agentic coding models Opus 4.7 and GPT 5.5, which were released in the prior week.

DeepSeek V4 beats all current open-source models in math, STEM, and coding, rivaling top closed-source models.

DeepSeek V4 is nearly state-of-the-art, almost as good as Opus 4.7 and GPT 5.5, and being nearly as good is sufficient for the vast majority of use cases.

On benchmarks like MMLU Pro, GPQA Diamond, and SWE-bench Verified, DeepSeek V4 is slightly behind Opus 4.7 and GPT 5.5 but still right up there, and it costs a fraction of the price.

The vast majority of use cases do not require absolute frontier-level intelligence, so DeepSeek's greater efficiency and lower cost are a real problem for the United States.

In a price vs. intelligence chart, GPT 5.5 and Opus 4.7 are the most expensive, while DeepSeek V4 Pro is slightly lower intelligence but much cheaper, and V4 Flash costs pennies per million tokens.

Since GPT-4 in May 2023, Chinese open-source models (Qwen, GLM4, DeepSeek R1) have repeatedly closed the gap with US models after each US advance, though they have always remained slightly behind so far.

US export controls prevent Nvidia from selling its top GPUs directly to China, but China compensates through algorithmic innovation and likely smuggling, enabling them to create frontier models with restricted hardware.

Nvidia CEO Jensen Huang has argued that China will develop its own AI chips anyway, so it is better to sell them American chips so their infrastructure is built on US technology.

DeepSeek V4's efficiency and open-source nature make it a big threat to the US economy because it will be very attractive to US companies and allies.

Anthropic published a report stating they have proof that top Chinese AI labs have been conducting distillation attacks on Claude to steal model data.

The US government confirmed that foreign entities primarily in China are running industrial-scale distillation campaigns using tens of thousands of proxies and jailbreaking techniques to steal American AI breakthroughs.

DeepSeek's alleged distillation attack involved only 150,000 exchanges, far fewer than other Chinese labs (Moonshot 3.4 million, Miniax 13 million), which is insufficient to explain their model quality and could simply be benchmark comparisons.

DeepSeek's own white paper acknowledges severe compute constraints, with very limited Pro service capacity, and expects price reductions after deploying 950 super nodes in the second half of the year.

GPT 5.5 costs $30 per million output tokens, and Opus 4.7 is similarly priced.

DeepSeek is open-source, can be fine-tuned and self-hosted, and costs a fraction of US models, giving enterprises more control and a much lower bill.

More and more US and allied enterprise companies will choose to build on top of Chinese open-source technology, creating a strategic dependence and a security risk if China cuts off access or changes architectures.

The US has trillions of dollars pouring into AI and infrastructure buildout, and if returns fail to materialize due to cheaper Chinese alternatives, the US economy could collapse, especially given the reliance on AI as the future of the economy.

If the world builds on Chinese AI models, China could dictate narratives and control what models are allowed to say, similar to how US-originated social media shaped global discourse.

The United States needs to push much harder on open-source AI because current frontier labs (except Google's smaller models) are not producing open-source models at the level of DeepSeek V4.

Even if US labs remain closed-source, they must drastically reduce costs and become much cheaper much more quickly so that enterprise customers find them cost-competitive.

Anthropic is currently doing everything wrong, while DeepSeek is doing everything right.