Cheaper intelligence isn’t created value
Xavier Vallée, Founder, Sapionic
The original post included a chart of the closed and open-weights model frontiers. It is on LinkedIn, linked below.
We read a lot here about AI subsidies. And it is true. The cost of tokens might go up by 10x or even 50x.
The tokens you buy from OpenAI and Anthropic are priced below cost, funded by venture money, waiting for the day the bill lands. That risk is real, and worth planning for.
But look at this graph.
The blue line is the closed frontier, the models you rent. The green line is the open-weights frontier, the models you can download and run yourself. For two years the green line trailed the blue one by many months, sometimes the best part of a year. A comfortable gap. A moat.
That gap has almost closed.
Kimi K3 landed last week at the top of the chart, level with Fable 5 and Opus 4.8.
K3 is the newest model from the Chinese lab Moonshot. It is built to be large but economical: rather than running the whole model on every request, it switches on only the small part relevant to the task, so it delivers frontier-grade answers without frontier-grade running costs.
It also has an unusually large working memory, able to read and reason over roughly a million tokens at once, enough to take in a full data room, contract set, or codebase in a single pass, and it handles images as well as text.
Moonshot says a couple of engineering advances make it about 2.5 times more efficient to run than its previous version, which is why it costs a fraction of the Western leaders.
The part that matters most for a business is ownership. K3 is the most capable “open-weights” model released so far, which means Moonshot will publish the model itself for download, not just rent access through their own service.
And that ownership is what changes the economics.
Hosting providers will run K3 and undercut one another to win customers, and the price of the same capability might fall, rather than increase.
But do not confuse cheaper intelligence with created value. A model you can run for pennies still does nothing until you point it at the right decision. Falling prices do not write the business case. They remove the last excuse for not building it.