Free on the device, and eighty per cent cheaper in the cloud
Apple opened its on-device model to developers on Monday. OpenAI cut the price of its reasoning model by 80 per cent yesterday. Two directions, one conclusion.
At its developer conference on 9 June 2025, Apple introduced a Foundation Models framework letting third-party apps use the on-device model behind Apple Intelligence — running locally, working offline, with no per-use charge. Yesterday, 10 June, OpenAI cut the price of its o3 reasoning model by eighty per cent and introduced a higher-compute o3-pro.
These point in opposite directions — onto the device, and further into the data centre — and they say the same thing about cost.
What each one actually means
On-device, at no marginal cost. Small tasks — summarising a note, tidying text, classifying something short — can now happen on a phone with no data leaving it and nothing metered. For those tasks, the privacy question and the cost question both disappear.
Eighty per cent off the expensive tier. The same model, the same day, at a fifth of the price. Not a smaller model, not a discount for volume — a straight repricing of the capability this blog spent December advising you to ration.
The conclusion for a buyer
Anything you decided against on cost is worth revisiting, and this is now the third time this blog has had to say that: in October after prompt caching, in February after the tiering, and now.
The pattern is consistent enough to plan around. Assume the price of any given AI capability falls by roughly half each year, and that some of that arrives in sudden steps rather than gradually. That means two things: do not sign long, and do not build a business case so tight that it only works at today’s prices — it will work better next year, not worse.
The thing that did not get cheaper
Everything in the third line of the May budget post. Cleaning the data, defining the process, checking the output for the first two months.
Compute has fallen roughly tenfold since this blog started in October. The internal work has not moved at all, and it is now the overwhelming majority of what an AI project costs a Swiss SME.
Which is a strange and useful conclusion: the cheapest part of your AI project is the AI.