Controlling AI Costs: Tokens, Models and Real-World Usage
Understand what drives AI bills, why they spike, and six ways to cut spend without cutting quality, from model choice to batching and prompt caching.
In the first month of using AI through an API, the cost looked so small that nobody paid attention. By month three, someone had connected AI to read every customer email, and someone else had built an agent that called the model over and over for a single task. When the bill arrived, finance asked a question nobody could answer: which work is this money going to, and is it worth it? If you are about to scale AI across your organisation, understanding the cost structure early saves you that surprise.
What is a token and why does it matter for cost?
A token is a small unit of text that a language model reads and writes. Most API-based AI services charge by the number of tokens, both input and output. OpenAI's help centre gives a rule of thumb that in English one token is about four characters, and notes that languages other than English often have a higher token-to-character ratio, which can make them more expensive. For Thai, measure real usage rather than guessing from word counts.
It also helps to know that output usually costs more than input. Anthropic's pricing page, for example, lists a higher output price than input price for every model. Answers that are longer than needed are a hidden cost.
Why AI bills spike without warning
- Context grows every turn: a long chat sends the whole history back to the model each time.
- Attaching whole documents: when only a few paragraphs are needed.
- Big models for easy work: such as sorting emails with a top-tier model.
- Looping agents: one task calls the model dozens of times, or gets stuck in a loop with no limit.
The OWASP Top 10 for LLM Applications 2025 also lists unbounded consumption as a key risk, which includes financial damage.
Six ways to cut costs without cutting quality
1. Measure cost per completed task, not cost per token
Anthropic's guidance on choosing models recommends comparing cost per completed task. A cheap model that fails often has to be rerun, which can cost more than a stronger model that finishes in one go.
2. Match model size to the job
Classification, form extraction and short summaries can often use a small model. Save large models for complex analysis. Test on real samples before deciding.
3. Trim context to what is needed
Use RAG to retrieve only relevant passages, summarise old chat history instead of resending it, and set a maximum answer length.
4. Use prompt caching
If your system prompt or reference documents stay the same each time, some providers offer prompt caching, where reading the cached part costs much less than the normal input price, as described in Anthropic's pricing documentation.
5. Batch work that does not need an instant answer
Anthropic states that its Batch API gives a 50 percent discount on both input and output for non-urgent processing, such as summarising a full day of customer reviews overnight.
6. Set budgets and alerts
Use separate API keys per project, set daily usage caps, and limit the maximum number of steps an agent can take.
An example scenario
Imagine a mid-sized insurance company using AI to summarise hundreds of customer emails a day. At first, every full email plus the entire contact history went to a large model. After the review, the team split the work into two layers: a small model sorts emails first, and only complex complaints go to the large model, with history trimmed to the latest relevant messages. The daily summary report moved to an overnight batch. The result is that the team now knows exactly where each baht goes and can decide which model tier each task deserves.
The cheapest AI is not the model with the lowest price per token. It is the system that gets the job done in one pass.
Watch out for
- Provider prices and offers change often. Check the latest pricing page before every budget calculation.
- Cutting costs until quality drops can cost more than you save. Always measure quality alongside cost.
- Caching and batching have technical conditions that differ by provider. Read the documentation carefully.
How to get started
- Pull last month's usage report, split by project or API key.
- Find the three tasks that use the most tokens.
- Try a smaller model on 20 to 50 real samples and compare quality.
- Turn on caching or move non-urgent work to batch.
- Set a monthly budget and review it every month.
If you are building multi-step agents, start with the basics in What is an AI agent. To build practical AI skills, browse courses at Plants.work Learn.