DeepSeek V4 Is Here and the Price Tag Should Terrify Every American AI Lab

# DeepSeek V4 Is Here and the Price Tag Should Terrify Every American AI Lab

On April 24, DeepSeek dropped the first preview of its V4 series. Two models: V4-Pro and V4-Flash. Both are Mixture of Experts with a 1 million token context window. Both are MIT licensed. And both are priced like DeepSeek is trying to start a pricing war they know nobody else can win.

Let's cut to the numbers, because they're genuinely ridiculous.

DeepSeek V4-Flash: $0.14 per million input tokens, $0.28 per million output tokens.

GPT-5.5: $5 per million input, $30 per million output.

That's not a discount. That's a different category of product. V4-Flash undercuts even OpenAI's own budget tier, GPT-5.4 Nano, which runs $0.20/$1.25. And V4-Pro — the big one, 1.6 trillion parameters, 49 billion active — costs $1.74/$3.48. Claude Opus 4.7 is $5/$25. GPT-5.5 is $5/$30. Simon Willison ran the comparison and the table reads like a pricing spreadsheet from an alternate timeline where American AI companies forgot how to optimize.

This Isn't Just Cheap. It's Efficient.

The pricing isn't a loss leader. DeepSeek isn't burning VC money to buy market share. They achieved this by building a model that is structurally cheaper to run.

From their technical notes: in a 1M-token context, V4-Pro uses only 27% of the single-token FLOPs and 10% of the KV cache size relative to their own previous model, V3.2. V4-Flash pushes that even further — just 10% of the FLOPs. They didn't just train a bigger model. They trained a model that knows how to stay out of its own way.

This is the part that should keep OpenAI and Anthropic executives awake. The US response to China's AI rise has been to throw capital at the problem — $100 billion for Stargate, trillion-dollar data center buildouts, export controls on Nvidia chips. DeepSeek's response has been to ask a simpler question: what if we just didn't need that much compute?

The Open Weights Are the Weapon

V4-Pro is 865GB on Hugging Face. V4-Flash is 160GB. Both are fully downloadable, fully modifiable, fully runnable on your own hardware if you've got it. MIT Technology Review noted that this makes V4 "a win for Chinese chipmakers" too — domestic AI hardware gets a flagship model optimized for it, without waiting for American permission.

But the geopolitical angle, while fun, isn't the most disruptive part. The most disruptive part is that any startup, any research lab, any hobbyist with a big enough server can now run frontier-class AI without paying a recurring tax to a San Francisco API provider. The moat of "we have the biggest model" just got drained.

What This Actually Means

I've been saying frontier models are commodities for a while now. DeepSeek V4 is the receipts. When the gap between "best closed model" and "best open model" shrinks to the point where the open one costs 6-10x less, the word "frontier" stops meaning "best" and starts meaning "most expensive."

This doesn't mean OpenAI or Anthropic are doomed. They have distribution, brand, enterprise contracts, and ecosystems. But it does mean the pricing power they enjoyed in 2024 is gone. GPT-5.5 can be marginally better than V4-Pro on some benchmarks, but is it $25 per million tokens better? For most use cases, the answer is no. And "marginally better but drastically more expensive" is not a winning long-term position in a commodity market.

The other thing this proves: context length is the new parameter count. Both V4 models handle 1M tokens. That's not a party trick — that's the ability to dump an entire codebase, a legal contract, or a research corpus into the prompt and ask questions across it. When your model is both cheaper and has a deeper memory window than the competition, the competition's marketing slides start looking desperate.

The Bigger Picture

DeepSeek isn't a magic underdog story. They're a well-funded Chinese AI lab with government support, talented engineers, and a clear mandate. They've also had setbacks — personnel departures, delayed launches, and scrutiny from both Washington and Beijing. V4 isn't going to shock the industry the way R1 did in January 2025. The shock value is gone. What's left is something more durable: a steady, relentless pressure on pricing that American labs have no obvious answer to.

You can't tariff your way out of a 10x efficiency gap. You can't embargo better engineering.

The AI infrastructure arms race assumes that more compute equals more capability. DeepKeep keeps proving that more clever equals more capability — and clever is a lot cheaper than compute. If that pattern holds, the trillion-dollar buildout of American AI infrastructure starts to look less like a strategic advantage and more like a very expensive bet that efficiency doesn't matter.

It matters. DeepSeek just priced the proof.

Simon Willison: DeepSeek V4 pricing breakdown · MIT Technology Review: Why DeepSeek's V4 Matters · DeepSeek on Hugging Face