The Cheapest Brain Wins

ChatGPT and Claude made intelligence a subscription. China is trying to make cheap AI a commodity.

Open three tabs. In the first, ChatGPT asks for $20 a month. In the second, Claude asks for the same. For a lot of writers, coders and analysts, paying for both stopped feeling extravagant a while ago – one is better at this, the other at that, and together they quietly become a $480-a-year thinking bill that nobody quite decided to sign up for.

Then DeepSeek opens in the third tab and asks a question the other two would rather you didn’t hear.

What if intelligence should cost almost nothing?

That is the actual Chinese challenge to OpenAI and Anthropic, and it is not the one everyone braced for. The expected challenge was “our model is smarter.” The real one is harder to answer: our model may be good enough, and it costs a fraction of yours.

The AI war has reached the checkout counter

The first phase of the AI race was easy to follow. Build the biggest model, hire the best researchers, buy the most chips, win the benchmark, ship the press release. The second phase is duller and probably matters more: how cheaply can you deliver intelligence to millions of people, thousands of businesses and a few billion devices?

For ordinary users, Western AI has settled into a familiar shape. ChatGPT Plus is $20 a month. Claude Pro is $20 billed monthly. There are pricier tiers for people who burn through usage. But you are not only buying answers – you are buying coding tools, research, file analysis, projects, integrations, and a steadily more polished place to do your work. The subscription is really a bundle, and the bundle is the point.

Developers get a different bill, itemised in a currency most people never see: tokens, the little chunks of text a model reads and writes. A million of them is about 750,000 words, so “a million in, a million out” is roughly the work of reading something longer than War and Peace and writing a fresh one straight back.

At today’s published prices, that job costs about $35 on OpenAI’s GPT-5.6 Sol. On Claude Sonnet 5, at its introductory rate – good through August 2026 – it is $12. On DeepSeek’s V4-Pro it is $1.30. Same job, same units. One end is a decent dinner; the other end is less than the tip you would leave on it. For that $1.30, DeepSeek will read your War and Peace and write you another one back.

These are not the same product, and the cheapest one is not automatically the best. But when the gap is not 20% but somewhere between nine and nearly thirty times, procurement stops reading the benchmark tables and starts reading the invoice.

DeepSeek is not doing magic. It is wasting less.

The lazy story is that Chinese AI is cheap because Chinese firms simply charge less and figure out the economics later. The more interesting story is that a lot of the price is engineering.

Start with the number that reset the whole conversation. DeepSeek says the final training run for its V3 model took 2.788 million GPU-hours – about $5.6M in rented computing time. That figure leaves out the research, the salaries and the runs that failed, so it is not the true all-in cost, and DeepSeek says as much. But the comparison still stings: models like GPT-4 have been reported to cost north of $100M to train. The lab did not spend its way to the frontier. It engineered its way there, and then priced accordingly.

The trick with a name is mixture of experts. Picture a hospital with hundreds of specialists. No patient sees all of them; each case is routed only to the few doctors it actually needs. DeepSeek’s V4-Pro is built the same way. It holds 1.6 trillion parameters, but only about 49 billion of them – roughly one part in thirty – switch on for any given word. The whole brain is there. Most of it just doesn’t clock in for every sentence.

Then there is the caching, the part that rarely makes the launch-day headlines. When a model re-reads text it has seen before – the same system prompt, the same document, the same codebase – DeepSeek can bill it at a cache rate that runs roughly fifty times cheaper than reading it fresh. The first pass costs money. The second pass is nearly free. For any software that leans on the same context all day, which is most software, that is where the bill quietly falls through the floor.

In plainer terms:

Western AI often looks like a bigger engine. Chinese AI increasingly looks like a better gearbox.

DeepSeek has also made leaving easy, which is its own kind of pricing strategy. Its API speaks the same format as OpenAI’s and Anthropic’s, so a developer can often swap the model behind an app without rebuilding the app. Cheap is a discount. Cheap plus a one-line migration is a crowbar.

This is bigger than one cheap chatbot

DeepSeek gets the headlines, but China’s AI strategy does not rest on DeepSeek alone. Alibaba’s Qwen family runs flagship models for the hard problems and startlingly cheap “Flash” models for the everyday high-volume grind, then knocks the bill down further with caching and 50%-off batch processing. The pattern is not one chatbot built to beat ChatGPT. It is a supermarket of intelligence.

The freshest proof landed in July 2026, and it did not come from DeepSeek. Moonshot AI, a Beijing lab backed by Alibaba and Tencent, released Kimi K3 – a 2.8-trillion-parameter model it calls the largest open-weight system yet built. On Arena’s frontend-coding leaderboard, where developers vote blind on which model builds the better interface, K3 took first place with 1,679 points – ahead of Anthropic’s Claude Fable 5 (1,631) and OpenAI’s GPT-5.6 Sol (1,618). One release earlier, its predecessor had sat in eighteenth.

Read the fine print and it is more measured. On the broad aggregate benchmarks, K3 still trails the very best American models, Fable 5 and GPT-5.6 Sol. What it beat, convincingly, was the tier just below – Claude Opus 4.8 and GPT-5.5 – and it did so as an open-weight model priced like Claude’s mid-tier Sonnet, at $3 / $15 per million tokens. The same trick as DeepSeek, one storey up: activate just 16 of its 896 experts per word, and a near-3-trillion-parameter brain stays cheap to run.

Markets skipped the fine print. By the next morning, AI and chip stocks were selling off across Asia; traders called it a “Kimi moment,” an echo of the DeepSeek shock that had erased roughly $600 billion from Nvidia in a single session. And all of it was built under three years of tightening American controls on the exact chips the model was supposed to need.

And here is the part that would have sounded absurd two years ago: some of these models are not rented at all. DeepSeek’s weights are public and MIT-licensed; Kimi K3’s are due out within weeks; much of Alibaba’s Qwen family is open too. You can download a model that competes with the frontier and run it on your own hardware, no subscription, no per-token meter running. The cheapest option, it turns out, is sometimes not a thing you buy. It is a thing you own.

The Western labs still carry themselves like premium software companies; the Chinese labs increasingly behave like industrial manufacturers, whose goal is not to be the best in any one room but to be in every room.

The buyers have already voted

None of this would matter if it were only a price list. What makes it a shift is that the customers have started moving – and the clearest record sits on OpenRouter, a neutral marketplace where developers route traffic across hundreds of models and pick whichever wins. It is one of the few places you can watch what people actually run, not what they say they admire.

On that platform, the American share of tokens has roughly halved in a year – from about 70% in mid-2025 to around 30% by mid-2026. Chinese models filled the gap. Look specifically at US firms, and the slice of their tokens going to Chinese models has stayed above 30% every week since February and peaked near 46%, up from an 11% average the year before. DeepSeek alone now moves more tokens than any other single provider on the platform.

The switches are not theoretical. One AI startup, Lindy, moved 100% of its traffic off Claude and onto DeepSeek, a change it says will save it millions. A fair caveat: OpenRouter is a routing marketplace, not the whole market, and it misses the traffic firms send straight to OpenAI or Anthropic. But it is the cleanest public read there is, and it has pointed one way for a year.

And then the state showed up

For a while this read like a story about companies. On July 17, 2026 – the day after Kimi K3 rattled American chip stocks – it became a story about the state. Xi Jinping took the stage at the World AI Conference in Shanghai and gave the keynote himself, the first time China’s leader has personally headlined the event. His subject was not any single model. It was the rules.

Xi announced a new body, the World Artificial Intelligence Cooperation Organization, to be headquartered in Shanghai and already signed by 29 countries, and cast China’s open-source models as a global public good – a counterweight, as Reuters read it, to American control of the field. He warned against any single country dominating AI and took a thinly veiled swipe at Washington’s export controls, cautioning against “overstretching the national security concept.” AI, he said, should be not a solo performance by one country but “a symphony of international cooperation.”

Read through the incentives, the design is elegant. The cheapest and most open models are Chinese; the developing world buys on price; and the country making those models is now offering to write the rules everyone else plugs into. Undercut on cost, win on volume, set the standards while the room is busy saving money. The American posture is the mirror image: its most capable model, GPT-5.6 Sol, is gated to roughly twenty vetted organisations. One side is rationing its best brain. The other is giving its away and building the head office.

Of course, cheap can get expensive

There is a reason people keep paying for the expensive names, and Kimi K3 handed over the evidence in its own launch numbers. The same model that topped the coding arena also watched its hallucination rate – how often it states something false with full confidence – climb from 39% to 51% against the version before it. More right answers, and more invented ones, in a single release. A model that costs 90% less but needs more checking, more supervision, or falls over when it matters is not actually cheaper. The token bill is just the number printed on the receipt.

The real cost of AI is the mistake nobody notices.

OpenAI and Anthropic still hold real advantages – polished products, familiar interfaces, deep tooling, and enterprise relationships that took years to build. China doesn’t have to erase any of that overnight. It only needs its models to get capable enough, available enough and cheap enough that a company stops reaching for the Western option by reflex. That reflex is the whole asset, and it is the thing under attack.

The first AI war asked who could build the smartest machine. The next one asks who can make intelligence so cheap that nobody hesitates before using it. China may not own the best brain in any given room.

But it is trying to decide what thinking should cost.