The cloud AI bill stopped being predictable
We are not saying the cloud costs more. We are saying something different and less comfortable: that in 2026 it stopped being something you can budget for. Here are the moves, with the public list price and the date we read it.
Four providers, one quarter, the same direction
Between March and August 2026 the four major artificial intelligence providers changed the way they charge. They did not all change in the same way, but they all moved towards the same place: less fixed price per person, more price for what gets consumed.
| Provider | What changed | From |
|---|---|---|
| Anthropic | Removed the bundled tokens from the Enterprise plan: you pay the seat (20 USD) plus consumption at API rates. The previous terms closed on 08/03/2026. | March 2026 |
| Microsoft | Raised Microsoft 365 list prices by between 5% and 33% depending on the edition (E3 from 36 to 39 USD, Business Basic from 6 to 7, F3 from 8 to 10). Copilot as a separate product stayed at 30 USD. | 01/07/2026 |
| OpenAI | Introduced a “Premium” seat in ChatGPT Business at 100 USD/month on the annual plan. The Standard seat remains. | 25/08/2026 |
| Gemini Enterprise moved to pay-as-you-go with no base fee, with configurable spending caps. | 26/08/2026 |
Public list prices in dollars, read on 30/08/2026. We do not convert them into local currency: the price you will be charged is the one on the list, in the currency of the list. Sources: Microsoft Licensing News · Google Cloud Blog · OpenAI's announcement on its official forum · The Register's coverage of the Anthropic change.
Why this matters more than the price
A high price can be negotiated, compared, put into the budget. A price that depends on usage cannot be budgeted for: it depends on how much your team works next year, and on technical decisions you do not make.
The case that illustrates it best appeared in Fortune on 07/08/2026: the company Pylon saw its Anthropic bill go from around 400,000 to 1.4 million dollars once it passed 150 seats — not because it used more, but because the entry terms had expired. That is the exact difference between a rise in consumption and a change of rules.
There is also a technical multiplier that almost nobody explains to the customer. When a task is routed to a “reasoning” model — the kind that thinks step by step before answering — consumption jumps. Gartner, quoted by The Register on 17/08/2026, estimates that it costs at least five times more to run than the same task on a normal model. And who decides when one or the other gets used is not you: it is the provider.
What changes with a machine in the office
Conectao Core inverts the structure of the spending; it does not remove it. You pay for the machine once, for the installation once, and for an annual maintenance plan you know in advance. Usage is free: one more query does not cost one more query, because the intelligence is already inside and the electricity of a desktop machine does not move a set of accounts.
There is no per-employee licence either. The software we install is professional open source, public and verifiable: adding the tenth person to the team does not open another subscription.
And it works with no internet. The connection is there for maintenance and updates, with your permission — not for working. In a hotel answering guests at eleven at night, that is not a technical detail.
When this is NOT the right answer
A commercial page that only lists advantages is no use for deciding. These are the cases in which we would tell you not to buy:
- Occasional, infrequent use. If it is going to be a handful of queries a month, the machine never pays for itself. A cloud subscription is cheaper and starts today.
- You need the best model in the world for one specific task. The frontier models in the cloud are still ahead (see the article on the eleven points). If your case lives inside that margin, the correct answer is the cloud.
- Nobody decides who sees what. If the company archive has no owner and no permissions, putting AI on top of it does not tidy anything: it amplifies the disorder. The archive comes first.
- The process will change within six months. Automating a procedure that is still being invented means paying twice. Stabilise it first, automate it afterwards.
The honest way to decide
Add up what you pay today in AI and software subscriptions, per person and per month, and multiply by twelve. Then ask yourself what happens if that number doubles without you doing anything differently. If the answer worries you, a machine is worth looking at. If it does not worry you, you probably do not need one yet — and we will tell you so in the demo anyway.