DeepSeek’s new V4-Flash model costs about three cents to clear a full pass of Artificial Analysis’s nine-benchmark Intelligence Index. Anthropic’s Claude Fable 5 costs $3.15 to do the same job. That’s not a rounding difference; that’s two orders of magnitude, and it’s why I don’t think Friday’s release is really about DeepSeek trying to top a leaderboard again.

DeepSeek hasn’t topped anything in a while. R1 was the model that triggered a global tech selloff back in early 2025, and since then Moonshot, MiniMax, Z.AI, and Alibaba have all crowded into the same lane, several of them scoring higher than V4-Flash’s 50 on the Intelligence Index (Moonshot’s Kimi K3 hits 57, good enough for fourth place overall, right behind Anthropic’s Opus 5 and Fable 5 and OpenAI’s GPT-5.6). What DeepSeek still does better than anyone is make the alternative embarrassingly cheap, and on the same Friday it shipped V4-Flash, MiniMax put out an open multimodal model that folds video generation, editing, and understanding into one system, and ByteDance stretched its Seedance video generator out to 30 seconds a clip. China’s AI labs apparently agreed to turn in their homework the same day, and Alibaba followed a few days later with a model Reuters says isn’t far behind Kimi K3 in scale either.

Here’s the part I find more interesting than any single benchmark: this pricing collapse has split Silicon Valley into two camps that don’t agree on almost anything else. Jensen Huang posted his first-ever tweet on July 24th, an open letter arguing that open-weight models expand the addressable market for AI chips, and Microsoft, Google, Meta, and even OpenAI’s Sam Altman signed on within days (“i want the US to win in AI both in open source and proprietary models,” Altman wrote, an odd thing to say if you also think China’s open models are a straightforward security threat). 179 startups wrote to the Trump administration asking it not to cut off their access to Chinese weights. Anthropic went the other direction entirely: Dario Amodei has been arguing publicly that Beijing could use these models to chase military superiority or tighten domestic surveillance, and that the responsible move is restriction, not embrace.

I get both arguments, and I don’t think either side is being cynical. Nvidia sells more silicon when more people can afford to run models on it, full stop, and that’s a real business interest talking. But Anthropic is the company that got its own Frontier deployment pulled by the US government during the Fable 5 shutdown, so its instinct toward restriction isn’t purely commercial either. This is the same compute-and-off-switch question I keep coming back to, just wearing a price tag instead of an export license this time, and export licenses haven’t exactly been airtight anyway: Taiwan is still arresting people for smuggling Nvidia chips, three arrests into that investigation with no export law to actually charge them under.

What actually decides this fight probably isn’t either company’s letter. It’s the number one researcher put on the other side of the ledger: banning Chinese open-weight models could cost American businesses something like $12 billion a year, estimated from OpenRouter usage data during one week in late July, because developers have quietly been routing simpler tasks to the cheap Chinese models to save money on inference bills that Goldman Sachs says fell from $2.07 to $1.67 per million tokens between June and July, a drop it partly credits to exactly this competition. That $12 billion estimate is rough by the researcher’s own admission, and Foundation Capital’s Jaya Gupta raised the scarier version of the same argument: a chunk of the AI infrastructure buildout is financed on debt that assumes sustained demand, and choking off the cheapest models on the market is one way to find out how much of that demand was price-sensitive all along.

Moonshot’s Kimi K3 is the model that actually spooked Washington into moving on this, both for beating benchmarks and for the intellectual-property allegations attached to it, and OpenAI answered the only way it credibly could: Altman cut GPT-5.6 Luna’s price 80 percent and Terra’s 20 percent the same week, which tells you a lot about who’s actually setting prices in this market right now, and it isn’t OpenAI. I’ve spent most of the last year watching this exact race play out from the model side. What changed this week is that the fight moved from benchmarks to balance sheets, and balance sheets are a much harder thing for either Washington or Anthropic to argue with.

Sources