DeepSeek released V4-Pro on April 24, 2026, and within 24 hours the discourse split into two camps: people who read the benchmarks and declared it overhyped, and people who read the pricing and went quiet. The first group is technically correct. The second group understands what is actually happening.
V4-Pro trails GPT-5.5 on SWE-bench Verified, Terminal-Bench 2.0, and the overall AI Intelligence Index. DeepSeek's own technical report admits a 3-6 month lag behind frontier models, which Carter James at Oplexa Insights called "rare honesty from an AI lab." That honesty is worth something. It also does not change the math. V4-Pro costs $3.48 per million output tokens. GPT-5.5 costs roughly $30. V4-Flash, the lighter variant, costs $0.28. These are not rounding differences; they are structural ones.
The Chip Story Is the Real Story
The benchmark debate is a distraction from the question that matters more: how did DeepSeek build a near-frontier model without Nvidia? V4 was trained on Huawei Ascend and Cambricon chips, deliberately excluding Nvidia access, and Huawei confirmed full support on its Ascend 950 at launch. The US export controls that were supposed to slow China's AI development have instead produced a Chinese AI stack that does not depend on American hardware at all. That is not a side effect. That is the outcome.
The efficiency gains are real and specific. V4's Mixture-of-Experts architecture activates only 49 billion of its 1.6 trillion parameters at inference time, runs 1-million-token contexts at 27% of the compute cost of its predecessor, and uses 10% of the KV cache. These are not marketing numbers; they are the reason inference is cheap enough to undercut every major US competitor. DeepSeek expects prices to fall further as Huawei scales Ascend production. US labs, meanwhile, are locked into Nvidia costs that are flat to rising.
Here is where I have to grant the skeptics something: V4 genuinely underperforms on the tasks that matter most for enterprise software development. GPT-5.5 leads on agentic coding benchmarks, and for a company building autonomous software pipelines, that gap is real. A model that is 8x cheaper but fails on 20% of your critical tasks is not a bargain.
But that framing assumes the buyer is a US enterprise with GPT-5.5 already in budget. The more consequential buyer is the one who could not afford frontier AI at $30 per million tokens: the Southeast Asian startup, the Brazilian fintech, the African health platform. For those buyers, V4-Pro at $3.48 is not a compromise. It is access they did not have before, running on infrastructure that does not route through American cloud providers or American export policy.
Who the Cost Structure Actually Serves
The MIT license on both V4-Pro and V4-Flash means any organization can deploy, modify, and redistribute the model. Combined with Ollama's announcement that V4-Flash is available on its cloud without requiring a local 160GB download, the friction to adoption outside the US just dropped significantly. DeepSeek is not competing for OpenAI's existing customers. It is building the market OpenAI priced out.
V4 beats GPT-5.5 on exactly one major benchmark: LiveCodeBench, at 93.5% versus roughly 82%. One win, one category. The rest of the scorecard favors OpenAI. But the inference economics scorecard is not close, and it compounds every quarter that Huawei scales while Nvidia supply stays constrained. The benchmark question was always the wrong one. The right question is which cost structure wins over 3 years, and DeepSeek's answer is already running on Ascend supernodes without a single Nvidia chip in the stack.
