DeepSeek V4 Is Here — And It's 50x Cheaper Than GPT-5.5

SCENE ONE. THE DROPFifteen months ago, a small Chinese AI lab you had never heard of released a model that wiped six hundred billion dollars off Nvidia's market cap in a single trading session. That lab was DeepSeek. That model was R1. And today, Friday, April 24, 2026, they did it again. DeepSeek dropped the preview of their V4 model family. Two models shipped at once. DeepSeek V4-Pro, with one point six trillion parameters. DeepSeek V4-Flash, with two hundred eighty four billion parameters. Both are Mixture of Experts. Both ship with a one million token context window as the DEFAULT mode of operation, not a bolt on feature. Both released under the MIT License. Both already sitting on Hugging Face where anyone with a GPU and a weekend can pull the weights and run them locally. The pricing is the part you need to SIT DOWN for. V4-Flash costs fourteen cents per million input tokens and twenty eight cents per million output tokens. V4-Pro costs one dollar seventy four in and three dollars forty eight out. Now compare that to what OpenAI announced yesterday. GPT-5.5 costs five dollars per million input tokens and thirty dollars per million output tokens. GPT-5.5 Pro costs thirty dollars in and one hundred eighty dollars out. DeepSeek V4-Pro is coming in at roughly ONE FIFTIETH of the price of GPT-5.5 Pro. Not one third. Not one tenth. One FIFTIETH. And it is open source. You can download it. You can run it on your own hardware. You do not need to send a single token of your data to an American cloud provider. That is the release. That is why everyone in the AI industry stopped what they were doing this morning when it dropped on Hugging Face. Now here is the first thing that makes this release genuinely different from everything DeepSeek has done before, and I want you to pay attention because most coverage is going to miss the significance. DeepSeek did not give Nvidia early access to the V4 weights. They did not give AMD early access either. The people who got the model first, for hardware optimization, were Huawei and Cambricon. Two Chinese chip companies. That is a REVERSAL of how the AI industry has worked for the last decade. Western chipmakers have always been the first to tune for new frontier models. That assumption has now been broken on purpose. The preview is explicitly tuned to run on Huawei Ascend chips. The CUDA lock in that has been Nvidia's moat for twenty years is being actively pried open by the second largest AI market on earth. Jensen Huang went on the Dwarkesh podcast two days ago and said this would be, quote, a horrible outcome for America. He is not wrong about the strategic picture. And here is the timing. The V4 launch landed less than twenty four hours after the White House accused China of industrial scale intellectual property theft from American AI labs. DeepSeek did not respond to the accusation. They just dropped a frontier grade open model at near zero cost and let the release speak for itself. SCENE TWO. THE NUMBERSLet me walk you through the benchmarks, because DeepSeek is claiming a lot, and some of the numbers actually hold up. On SWE-bench Verified, which measures real world software engineering task completion pulled from actual GitHub repositories, DeepSeek V4-Pro scored eighty point six percent. Anthropic's Claude Opus 4.6 scored eighty point eight percent on the same benchmark. That is a gap of zero point two percentage points. Statistical noise. DeepSeek V4-Pro is functionally tied with Claude Opus 4.6 on the benchmark that serious software engineers actually care about. On Terminal-Bench 2.0, DeepSeek V4-Pro scored sixty seven point nine percent. Claude scored sixty five point four percent. DeepSeek is winning this one by two and a half points. On LiveCodeBench, DeepSeek scored ninety three point five percent. Claude scored eighty eight point eight percent. DeepSeek wins by nearly five points. On Codeforces, DeepSeek V4-Pro put up a competitive programming rating of three thousand two hundred and six. That is grandmaster territory. Now the honest caveat. On certain general reasoning and frontier intelligence benchmarks, DeepSeek V4-Pro trails GPT-5.4 and Gemini 3.1 Pro by a margin that works out to roughly three to six months of development lag behind the absolute bleeding edge. So DeepSeek is NOT claiming the crown. They are claiming parity with Anthropic on coding, and near parity with OpenAI and Google on general intelligence, at a price point that is not even in the same galaxy. Let me redo the pricing math because it is worth hearing twice. V4-Flash. Fourteen cents input. Twenty eight cents output. Per MILLION tokens. To put that in context, you could run the complete works of Shakespeare through V4-Flash for about forty five cents. V4-Pro. One dollar seventy four input. Three dollars forty eight output. Per million tokens. Claude Opus 4.6, for comparison, costs fifteen dollars per million input tokens and seventy five dollars per million output tokens. That is roughly a nine times price gap for functionally identical SWE-bench performance. V4-Flash is now, by a clear margin, the cheapest small model on the market. It beats GPT-5.4 Nano. V4-Pro is the cheapest frontier class model on the market. And both are open weights. There is one more specification that is going to matter more than most people realize. The one million token context window is NATIVE to V4. It is not a retrieval trick. It is not a sliding window. It is not augmentation. The model was trained from the ground up to reason across one million tokens of context as its default mode. Other models support long context, but most of them degrade in ways V4 reportedly does not. If that holds up in independent testing, and we are still waiting for independent testing, then V4 is the first open model family designed around million token context as the DEFAULT rather than as an edge case. The Hacker News thread on the release is already a thousand comments deep. The top comment says, quote, the Chinese ecosystem has delivered a complete AI stack. Like it or not, that is big news. That is a fair read. Hugging Face downloads are already in the tens of thousands. The Zhipu AI and MiniMax stocks fell nine and seven percent respectively in Asian trading overnight. Their own Chinese competitors just got undercut by the OPEN SOURCE release of a model that beats them at a fraction of the cost. SCENE THREE. THE REAL STORYHere is what I think most people are going to miss, because the coverage is going to focus on the numbers, and the story is not really the numbers. The story is the chip stack. For twenty years, American AI leadership has rested on two pillars. First, the best models in the world were built in American labs. Second, the hardware they ran on was American hardware. Nvidia, AMD, with CUDA as the software layer that tied everything together. If you wanted frontier intelligence, you came through those companies and paid that tax. DeepSeek V4 pulls on both pillars at once. The model is near frontier. The model is open. The model is tuned for Chinese chips from day one. The model is the cheapest of its class by a factor that makes the numbers hard to rationalize for any company buying from the American stack. That is not a product release. That is an industrial policy statement delivered in the form of model weights. Now the bad faith reading, because I want to give you the full picture. Anthropic accused DeepSeek earlier this year of running thousands of fraudulent Claude accounts to generate training data through distillation. OpenAI has made similar noises about their own APIs. The gap between DeepSeek's compute budget and its output quality has always been smaller than it probably should be. The distillation allegations are UNPROVEN, but they are also not crazy, and they sit in the background of every DeepSeek release including this one. There is also the Huawei question. Some analysts believe DeepSeek was pushed by the Chinese government into the Huawei optimization story as a geopolitical signal. Others believe it happened organically because Huawei Ascend chips are simply what Chinese AI labs can reliably buy now under American export controls. Both readings are plausible. Neither has been confirmed. I mention the conspiracy version because it is circulating widely, not because I endorse it. Jensen Huang's reaction on the Dwarkesh podcast deserves to be played in full. He said a DeepSeek model running at production scale on Huawei silicon would be, quote, a horrible outcome for America. That is the CEO of the single most important chip company in the world saying the moat he built is being crossed in real time. And the TIMING of the release. Less than twenty four hours after the White House accused China of industrial scale IP theft, DeepSeek pushed a one point six trillion parameter open weight model to Hugging Face, and the message is not subtle. It says, we do not need your labs. We do not need your chips. We do not need your licenses. And we can ship at one fiftieth of your price. So where does this leave things. If you are a developer, V4 is probably the highest ROI model you can run today, and you should be testing it by Monday. If you are an enterprise buyer, you now have an open source option at near frontier quality, and that is going to change procurement negotiations across the industry. If you are OpenAI or Anthropic, you have a pricing problem that is not going away, and a narrative problem that is going to get worse every time DeepSeek ships. If you are Nvidia, you are watching the second largest AI market on earth build a hardware stack that does not need you. The R1 moment was a warning shot. V4 is the response. Stay sharp. Jane Sterling, Sterling Intelligence.