AI costs are not only about the price per token.
Cloud AI is cheap for ordinary use. The bill becomes hard to forecast when AI turns into infrastructure, because the customer sets none of its three variables: the price, the volume and the model.
For chat, drafting and occasional analysis, a usage-based API costs less than dedicated hardware. The company pays only for what it uses, and no server sits idle. The figures below show how small that bill is, and what moves it.
The price: the vendor sets it
| Date | What Anthropic announced | Input and output | Source |
|---|---|---|---|
| 30 June 2026 | Claude Sonnet 5 launches at introductory pricing | $2 and $10 | Anthropic |
| 30 June 2026 | The same announcement schedules a rise for 1 September | $3 and $15 | Anthropic |
| 10 August 2026 | The post is edited: the introductory price “is now permanent” | $2 and $10 | Anthropic |
Anthropic gave notice weeks ahead and then kept the lower price. The budget line still faced a 50 % swing, decided twice in six weeks, and the customer made neither decision. Cloud prices also fall. The customer does not set those either.
The volume: agents multiply it
| How the team works | Tokens a month | Cloud bill a month |
|---|---|---|
| Chat and drafting: about 95,000 tokens per person per working day, roughly 24 turns of 4,000 tokens | 100 million | 80 x $2 + 20 x $10 = USD 360, about EUR 330 |
| The same team with agents, if the tokens grow tenfold (our assumption) | ten times as many | USD 3,600, about EUR 3,300 |
Agents review documents, draft responses, check their own work and produce another version, so the tokens grow and the head count does not. The factor of ten is our assumption, not a measurement. With it, the line moves from EUR 330 to EUR 3,300, and nobody signs a purchase order.
The model: the vendor retires it
OpenAI's API deprecations page lists the shutdown of the chatgpt-4o-latest model on 17 February 2026, with notice given on 18 November 2025.
The replacement is often better, faster or cheaper. It may also behave differently, and then part of the work built around the old model is done again: the prompts tuned to it, the automated tests, the compliance approvals, the expected output formats and the downstream integrations. The cost is the API bill plus that rework.
The threshold: about 165,000 tokens per user per day
A fixed monthly fee wins only above a volume, and the volume can be calculated. Bastion's Reasoning appliance costs EUR 5,000 a month. Claude Fable 5.2, Anthropic's newest frontier model, is listed at $8 and $40 per million tokens on 17 November 2026. At the same 80 % input share, 100 million tokens cost 80 x $8 + 20 x $40 = USD 1,440, about EUR 1,310 at the same ECB rate. The two bills are equal at 5,000 / 1,310 x 100 million, about 382 million tokens a month.
The appliance serves up to 110 people with thinking off, at 8k tokens of context, at 30 tokens per second each (p90), with an fp8 KV cache (based on a test on a rented 600 W card with the same chip, before our first box is built). 382 million tokens across 110 people and 21 working days is about 165,000 tokens per person per day. That is document processing and agent work, not conversation. Below it, the cloud is cheaper.
What finance should estimate before choosing
Estimate the workload after adoption succeeds, not the workload of today: the users, the automated workflows, the documents processed, the tokens per task and the rate of growth. Then set the variable cloud bill beside the fixed fee. A small or uncertain workload gains from the elasticity of the cloud. A large and growing one gains from a cost that finance can forecast.
With Bastion
What Bastion changes
Bastion is a private AI system delivered as one sealed box that runs inside your building. One monthly fee covers the hardware, the model, the hardened operating system and support, and nothing your team types leaves the building.
Bastion is one monthly fee for an appliance that serves the team. The fee covers the hardware, the model refreshes and the support, and it does not move when usage does. The model changes when the organisation accepts the update, not when a provider retires one.
A team that uses AI for ordinary chat will pay less in the cloud, and it will stay that way. The other two reasons on this site, where the data goes and who holds the switch, stand on their own.
Questions this article answers
- How much does cloud AI cost a company per month?
- A 50-person team using 100 million tokens a month pays about EUR 330 at Claude Sonnet 5’s June 2026 prices. When the team starts agents, use can grow tenfold, and the bill reaches about EUR 3,300.
- Can a cloud AI price change after we sign?
- Yes. Anthropic launched Claude Sonnet 5 at introductory prices on 30 June 2026, scheduled a rise for 1 September, and on 10 August made the lower price permanent. The customer made neither decision.
- When is a local AI appliance cheaper than cloud?
- At roughly 165,000 tokens per user per working day, in our calculation for one Reasoning appliance against a frontier cloud model at its list price of 17 November 2026. Below that, cloud is very hard to beat.
On the record
- 1
- 2
- 3
- 4