Insights
Nobody told your finance director the floor fell out of AI
China's open models now undercut Claude and OpenAI by up to fifty times. Here is what that actually means for the bill you are paying, and the catch almost nobody mentions.
If your business pays to use AI, you are almost certainly paying more than you need to, and the gap is no longer small.
For the past year, running a top-tier American model like Claude Opus or OpenAI's flagship has cost around five dollars per million words of input and twenty-five dollars per million out. Everyone treated that as simply the price of good AI. A new generation of Chinese models has now broken that price, doing the same work for a fraction of the cost. On the majority of real business tasks, you would struggle to tell the output apart.
This is the biggest shift in AI economics this year, and most companies have not adjusted for it. Here is what changed, what it means for the bill you are paying, and the one catch that genuinely matters.
The numbers
At the time of writing, DeepSeek's current workhorse model charges about fourteen cents per million tokens in and twenty-eight cents out. Not five dollars. Fourteen cents. That is the same unit of work the American labs bill at thirty-six times the price, and on the output side you can get to fifty-odd times cheaper if you pick your comparison. Alibaba's Qwen line goes lower still on its fast tiers. Moonshot's Kimi and Zhipu's GLM are in the same bracket.
These are not toys. They are not the discount own-brand version that falls apart the moment you ask it something hard. On the benchmarks that people actually argue about, the gap between the best Chinese open model and the best American closed one has narrowed to the point where, for the overwhelming majority of real business tasks, you would struggle to tell them apart in a blind test. The performance is comparable. The price is not remotely comparable.
Enterprise buyers have quietly worked this out. By 2026, American firms were routing close to half of their tokens on some platforms through Chinese open models. Not out of some ideological love of the thing. Out of arithmetic.
What it did to the companies charging you
You can tell how frightened a company is by how fast it cuts its own prices. Anthropic, the maker of Claude, dropped its rates by 67 per cent in a single announcement. OpenAI is reported to be weighing "drastic" reductions of its own. When two companies burning through a combined sixty-odd billion dollars a year, neither of them profitable, suddenly decide your bill should be a third of what it was last month, they are not doing it out of generosity. They are doing it because a Chinese lab set the floor and everyone else is now falling towards it.
Which is wonderful news for you, the person paying the bill. The single most useful thing you can understand about AI pricing in 2026 is that it is no longer a fixed cost. It is a choice, and most businesses are still making the expensive one by default because nobody has sat down and told them otherwise.
The catch
Here is the part the breathless "China won" coverage skips, and it is the part that actually matters if you run a real company with real data.
If you send your prompts to a model hosted in China, your data goes to China, and Chinese law travels with it. That is not a flag-waving point, it is a contractual one. A Chinese AI provider is subject to Chinese legal obligations regardless of what its privacy policy says, where its servers sit, or how reassuring the marketing is. For a lot of businesses, your client list, your case files, your unreleased numbers, that is simply a non-starter, and no discount makes it acceptable.
So if the story ended there, it would be a tease: incredible prices you are not allowed to use.
But it does not end there, and this is the bit worth paying attention to. The best Chinese models are not just cheap, they are open weight. DeepSeek ships under an MIT licence, Qwen under Apache 2.0. In plain English: you can download the actual model and run it on hardware you control, on Western infrastructure, in your own cloud account, with nothing phoning home to anyone. You keep the near-frontier capability. You keep the rock-bottom running cost. And your data never leaves the building.
That is the move. Not "should we use the Chinese API," which for most firms is the wrong question with an easy no. The right question is "which of these open models can we host ourselves, so the savings are real and the data risk is zero." The data risk was never set by who trained the model. It is set by where you run it.
What this means for you
Strip away the geopolitics and the benchmark noise and it comes down to three things.
First, if you are defaulting every task to the most expensive frontier model, you are almost certainly overpaying, and by a lot. A great deal of what businesses use AI for is bulk, unglamorous work: sorting, drafting, summarising, extracting. That work does not need the Rolls-Royce model. It needs something competent and cheap, and there has never been more of that on the shelf.
Second, the right setup is almost never one model for everything. It is a tiered one: a cheap, often self-hosted open model doing the heavy volume, and an expensive frontier model reserved for the genuinely hard five per cent where it earns its keep. Done properly, that alone can take a monthly AI bill down by the sort of percentage that makes a finance director briefly emotional.
Third, none of this is stable, and you should not want it to be. Governments are openly debating whether to ban Chinese models altogether, and the open-weight nature of them means such a ban would be almost impossible to enforce, which tells you how fluid the ground is. The specific cheapest model this quarter will not be the cheapest model next quarter. The winning strategy is not picking a favourite. It is building your systems so that swapping the engine underneath is a Tuesday afternoon, not a rebuild.
That last part is the actual work, and it is most of what we do. The models are a commodity now, and the commodity keeps getting cheaper. The value was never in knowing which one is best this week. It is in building a business that can quietly take the cheapest capable option every time, without you having to think about it, and without your data ending up somewhere it should not be.
If your AI bill has quietly become one of your larger invoices and nobody in the building can tell you why, that is usually a sign it is fixable. [Tell us what you are running](/enquire/), and we will tell you straight whether it is.