Models & Vendors
How does AI model pricing work?
The basic unit: what is a token?
Language models do not read and write in words but in tokens - small chunks of text. On average a token is about four characters, or roughly three quarters of an English word; longer languages and compound words produce more tokens per word. Every request and every response is broken into tokens and counted.
- Token = the smallest billing unit in usage-based AI pricing.
- Rule of thumb: ~1,000 tokens equal roughly 750 English words; somewhat fewer in German.
- Not just text: images, audio and attachments are also converted into token equivalents.
App subscription or API: two billing worlds
For end users there is usually a chat app (such as ChatGPT, Claude or Gemini) with a fixed monthly subscription. You pay a flat fee and use the model within fair-use limits, regardless of how many tokens are consumed internally. For developers and products there is the API: you pay per token actually used, but you can integrate the model into your own software.
- Subscription (app): predictable flat price, no setup, ideal for personal and occasional use.
- API: variable cost by usage, full automation, ideal for products and high volumes.
- Business and enterprise tiers blend both and add administration, access control and data-protection commitments.
- Some providers offer per-message billing or prepaid-credit models as a middle ground.
Why input and output tokens cost differently
In the API, input tokens (your prompt plus supplied context) and output tokens (the generated answer) are priced separately. Output tokens are almost always more expensive, because generating text is more compute-intensive than reading it. Crucial for cost control: the entire context you send - system instruction, chat history, uploaded documents - counts again as input on every request.
- Context window = the maximum tokens a model handles per request; larger windows allow more context, but every token you send costs.
- Long chat threads get expensive because the history is resent as input with every new message.
- Shorter, sharper prompts and selective context directly cut costs.
Discounts, caching and hidden cost drivers
Providers offer mechanisms to lower costs - and factors that quietly raise them. Prompt caching charges recurring context at a reduced rate, and batch processing discounts non-urgent bulk requests. Conversely, newer reasoning models may generate extra, often invisible thinking tokens internally, which are also billed as output.
- Prompt caching: repeatedly used context (e.g. a long instruction) is billed at a lower rate.
- Batch/async mode: a discount for tasks that need not finish in real time.
- Reasoning tokens: the internal thinking steps of advanced models silently raise output tokens.
- Model choice is the biggest lever: smaller models often cost many times less per token than flagship models.
Estimating AI costs: a simple mental model
For a rough API estimate, one chain suffices: estimated input tokens plus estimated output tokens per request, times requests per month, times the token prices published by the provider. Calculate input and output separately, since the rates differ. For apps, instead compare the subscription price against the expected value per active user.
- API formula: (input tokens + output tokens) x requests x token price = approximate monthly cost.
- Start with a smaller model and only move to a larger one when the need is proven.
- Set output limits, usage quotas and cost alerts before going into production.
- Never quote current figures from memory - always check the provider's pricing page.
Swiss perspective: currency, data protection and governance
Most major AI providers bill in US dollars; Swiss budgets must add the exchange rate and, depending on the case, VAT, so final CHF costs can fluctuate. Beyond price, processing matters: putting personal data into prompts falls under the revised Data Protection Act (revDSG/nFADP) and the oversight of the FDPIC (EDOEB). The EU AI Act applies extraterritorially and can affect Swiss providers serving the EU market.
- Budget in CHF with a buffer for exchange-rate swings and possible taxes.
- Clarify data location, data-processing terms and whether your inputs are used for training.
- Sovereign options such as the open Swiss model Apertus (ETH/EPFL) can strengthen data protection and cost control.
Frequently asked questions
What does a token cost?
No single figure is trustworthy, because token prices vary strongly by model, by provider and between input and output, and they change frequently. Prices are usually quoted per million tokens. For current figures, always check the provider's official pricing page.
Is a subscription or the API cheaper?
It depends on usage patterns. For regular personal use through a chat interface, the flat subscription is usually more predictable and cheaper. For automated, software-integrated or highly variable use, the usage-based API is often more economical - you pay only for what you consume.
Why are output tokens more expensive than input tokens?
Because generating text is more compute-intensive than reading it. The model computes each output token one after another, whereas the input context is processed in one pass. That is why providers usually list a higher price per output token.
What is a context window and does it affect price?
The context window is the maximum number of tokens a model considers at once per request. A larger window allows longer documents and histories, but it only costs when you actually fill it: you pay for tokens used, not for the window size itself.
What are reasoning tokens?
Newer models with strong reasoning produce internal thinking steps before the actual answer. These reasoning tokens are often invisible but are billed like output tokens. On complex tasks this can raise costs noticeably - a reason to deliberately match model to task.
How do I budget AI costs for a business?
Estimate input and output tokens per typical operation, multiply by expected volume and by current provider prices. Start with a smaller model, set cost limits and alerts, and budget in CHF with an exchange-rate buffer. Also check revDSG data-protection requirements before processing personal data.
Key terms in the glossary
Practical AI for your business
From idea to implementation – we show you what is concretely possible in your case.
Request a demo