What determines the cost of AI?

Everything is counted in tokens — fragments of words the model reads and produces. The longer the exchange (long prompt, large document, long response), the more tokens it consumes, and the more it costs. More powerful models also cost more per token than faster, lighter models.

What are the ways to pay for AI?

Four main models: free (limited, and your data may be used for training); personal subscription (a fixed monthly flat rate, often around twenty euros); pay-as-you-go API (you pay per token consumed, ideal for automation); and enterprise (per-user/seat billing with guarantees). Prices change fast — think in orders of magnitude, not fixed figures.

Subscription or API: which one to choose?

For someone who chats with AI every day, a fixed-rate subscription is the simplest and most predictable option. For automation, connecting AI to tools, or running agents, the API (pay-as-you-go) is the right choice.

How do you keep the bill under control?

Set spending caps on the API, pick the right model for the task (a fast model for volume, the powerful one only when it's justified), and constrain your agents (budget, frequency) to prevent runaway loops.

How much does AI cost: subscriptions, API, and keeping the bill under control

1. What drives the cost: tokens

AI doesn't bill "per question". Everything is measured in tokens — word fragments the model reads (your prompt, the documents you feed it) and produces (its response). The longer the exchange, the more it consumes, and the more it costs. Pasting an 80-page PDF "costs" more than a one-line question.

The second lever is your choice of model. A powerful model (think Opus) costs more per token than a fast one (think Haiku). That's the whole point of picking the right tool for the job — no need for a sledgehammer to hang a picture.

2. The ways to pay

Four main models, from simplest to most professional. Prices move fast: think in orders of magnitude, not fixed figures.

Free €0 For testing. Limited (volume, models), and your data may be used for training. Not for anything sensitive.

Personal subscription ~ flat rate /month A fixed amount (often around twenty euros). Use it as much as you want, predictable bill. The default choice for individuals.

Pay-as-you-go API per token You pay for what you consume. Essential for automation, connecting tools, and running agents.

Enterprise per seat Per-user billing with guarantees (confidentiality, administration, support). For teams.

Orders of magnitude, not a quote: pricing shifts with each new version. The logic stays the same.

3. Subscription or API: which one

This is the real practical question, and the answer depends on what you're doing:

Subscription

You're chatting

Fixed monthly flat rate
Predictable bill, zero setup
Ideal for everyday individual use

API

You're automating

Pay per token consumed
For connecting tools, running agents
More powerful, needs guardrails (caps)

Daily chat → subscription. Running automations → API. Many people end up using both.

4. Keeping the bill under control

The API can feel scary ("what if it spirals?"), but it's very easy to keep in check:

Set spending caps. Most API platforms let you set a hard monthly limit. Do it from day one.
Pick the right model. A fast model for volume and simple tasks; the powerful model only when the difficulty warrants it.
Constrain your agents. Budget, frequency, stopping criteria — otherwise an agent can loop and burn tokens for nothing (I cover this in the AI Agent guide).

For context: my daily HiFi monitoring runs on the API and costs next to nothing per month — because it's scoped (limited sources, once a day, right-sized model). Cost only becomes a problem when it's left unconstrained.

How much does AI cost: subscriptions, API, and a controlled bill