As you slowly move your cursor to the billing page, your heart trembles in fear. Pure angst races through your nervous system, as the tremors worsen. The page switches to blank, you don't want to see the result, but it is already set in motion. As the buffer slowly evaporates, your despair grows. Like a sledgehammer hitting a brick wall, your mind crumbles at the number on the screen.
$32,424.39 in API calls.
Image of your dream car or brand new iPhone fade away and make way for payment installments. It has happened again: You forgot to set limits in your OpenClaw account.
Does this feeling sound familiar? You are not alone! Many users of agentic platforms forget that using AI is fundamentally a business model, and as such costs money. Today we are exploring this phenomenon a bit deeper, looking at the how, why and things you can do to keep your savings intact.
First things first, what are agentic platforms? The concept consists of agents working together to achieve a goal without human intervention. An agent is basically a standalone AI model that has the ability to call certain tools. In the background this is nothing more than the agent spitting out JSON which outlines which function the model needs. In practice, it allows the agent to for example browse the web, build code files or communicate with other models. Agentic platforms use these agents in tandem with each other. By giving every agent a distinct system prompt with specific personalities, tool calling options and context, the system simulates a real life work situation. This is best explained by an example:
Let's say that you want to generate a report of this quarter's financials. You would talk to your accountant, which in turn works together with data specialist to obtain the data. This team also talks to the IT department, which builds a custom report generator. The end result: a nice-looking financial report.
Now we play the same scenario in an agentic framework. The initial AI thinks about which agent it needs to fulfill the request, and calls the accountant agent, the data specialist agent and the programmer agent. Multiple back and forths between the agents eventually result in the same end result: a nice-looking financial report.
The good part is, that in theory you can fully automate workflows via agentic platforms. The bad part however, is usually underestimated.
First of all, developing workflows is a discipline in itself. Many decades of business management have not yet resulted in the "perfect" workflow, and whilst you might guess that with agents this is easier, it most certainly is not. Without delving too deep into process management, you have to understand that achieving maximum efficiency is not a copy-paste matter and is heavily dependent on what it is you want to automate.
Secondly, agents will only pursue the goal that is given. Humans may come into office, start working and whilst ticking away develop other points of action. This iterative thinking is scarcely replicated in agentic systems, due to the fact that agents are designed to execute tasks within the context they operate. They will not, unless explicitly told, open the newsletter of today. It is not to say that this is not something that may happen later down the line, but many business decisions are influenced by someone's character and experiences, rather than a data point.
Lastly, and this is the big one, agentic systems are costly. And we mean, really costly.
One of the largest agentic platforms is called OpenClaw. And no, we are not slandering OpenClaw on pricing, their product is really awesome and we highly recommend experimenting with it, but we will break down why you need to be cautious. OpenClaw works by allowing you to set up different models for different agents. The reason you do this is to achieve higher quality output. For example, your "programmer" agent is best fitted with a high tier coding model such as Claude Fable 5, whilst a more strategic agent would flourish with Qwen3.5. Because most of us aren't able to run high tier models locally, we have to make due with API calls. And this is exactly the balancing act a lot of users fail. Ideally you want to have the highest quality output without using millions of tokens per execution. Because if you were to use the latest flagship models, you can easily accumulate a huge bill if you are not careful.
And by "not careful" we essentially mean "I don't know what I am doing, period". With these agentic workflows, it is quite easy for agents to get stuck in a loop, continuously calling the API, racking up those costs. And usually, the agents don't sleep, meaning this happens 24/7. For reference, owner and creator of OpenClaw Peter Steinberger posted a screenshot showing the API bill of OpenAI in May. $1.3M charged in 30 days. You can image that with the latest increase in token costs, you need to watch your bills.
Now that you understand the how and the why, let us dive into mitigating the costs. A good reminder here is that in essence, most older AI models are perfectly capable of executing tasks, especially in an agentic workflow. This already reduces costs significantly, as older models usually have a lower token cost.
Another thing to look out for is using smaller models for simple tasks, and larger models for harder tasks. OpenClaw for example allows you to run local models in tandem with cloud. Local models are usually smaller and perfectly suited for those simple or high volume tasks, and the only cost you incur is the energy bill of your PC. You can also choose small cloud models, such as Claude Haiku 3.5, which runs at $0.0001 per conversation. For our coding friends, Cursor has "auto" mode, which routes the cheapest and best suited model based upon the given task. Our development team manages to churn out 3-5 large scale Proof of Concepts per month this way, for a staggering $20 in costs.
You may also critically look at how much context you want to give each agent. Fully using a 256K context window might not yield better results compared to only using 125K tokens, but the costs are double. You may also opt to use a technique called Prompt Caching, which some models support. Essentially, this allows models to reuse system prompts and context throughout multiple steps without using the context window.
It is also good to preemptively set API spending limits. You'd rather have the system fail than to wake up with a large bill. Make sure those limits are set on the agentic platform as well as on the model provider, as bugs may bypass the agentic software. On top of spending limits, also set limits on how many times the system loops through itself. This way you make sure that you don't get infinite loops.
We advise to always check the documentation of the platforms themselves. Here is our list of web pages that display API cost estimates per agentic platform:
Take note that these are Bring-your-own-key products. Your workflow may be more suited for a single-pay model, where API costs are included.
Now that you are armed with a better understanding of agentic platform costs, you can safely dream again about that dream car. For now, we are happy to get into contact about agentic systems. At Initium, we build these systems from the ground up, perfectly fitted for your use-case. You know where to find us! Until next time!