The architecture almost nobody prices out first
An agent is not the model. An agent is a small program that decides what to do, calls a model when it needs to think, and acts on the answer. The program itself is light. It waits, it makes a request, it parses a response, it moves on. That part runs comfortably on a modest always-on server with a single CPU core and a gigabyte of memory.
The heavy part, the model, is a hosted API call. You send text, you get text back, and the GPU doing the work is someone else's problem. Your server never loads a model, never needs a GPU, never spikes past the small steady load of a program making network requests.
Split the job this way and the cost picture changes completely. You are no longer paying for inference hardware. You are paying for a small VM to keep your agent alive, and the inference comes from a provider's free tier. That is the whole trick, and it is genuinely underused.
There is a lot of free inference right now
The "make a project first, pay later" era for LLM APIs is real. A community-maintained list called awesome-free-llm-apis tracks providers with permanent free tiers, and the roster is not thin. Google's Gemini has a free tier. Mistral, Cohere, and Groq offer free access without a credit card. Fast-inference providers like Cerebras and SambaNova have free tiers. Aggregators like OpenRouter expose a couple dozen free models behind one endpoint. Most of these speak the OpenAI SDK format, so your code targets one shape and swaps providers by changing a base URL and a key.
For a builder, this means you can stand up an agent that reasons, summarizes, classifies, or answers, and pay nothing for the reasoning itself.
Where free stops being free
An opinion piece that stops here would be selling you something. So here is the honest part.
Free tiers are rate-limited, and the limits are built for one user, not a userbase. A typical free tier gives you something like a few dozen requests per minute and a few hundred to a few thousand requests per day. That is plenty for a personal agent, a Telegram bot you and some friends use, a nightly job that processes a queue, a scraper that summarizes what it finds. It is not enough for a product with real traffic. The moment you have hundreds of users hitting the model, you have outgrown the free tier and you should be paying for it.
Know which side of that line your project sits on. If it is a tool for you, free tiers are a gift. If it is a service for others, treat the free tier as a prototype and budget for the paid one before launch.
The second thing to know: several providers reserve the right to use free-tier prompts to improve their models. For a throwaway side project that is a non-issue. For anything touching data you would not want in a training set, read the provider's terms and pick one that does not train on your traffic, or pay for a tier that contractually does not. This is a real choice, not a formality, and only you know how much it matters for what you are building.
What you actually need
Strip it down and the shopping list for a cheap, always-on agent is short. You need a small server that stays up so your agent is there when its trigger fires. You need an API key from a provider whose free tier fits your volume. You need a bit of code that ties them together. That is the entire stack.
The server is the part people overspend on, because they assume "AI project" means "GPU." It does not. A single CPU core and a gigabyte of memory will happily run an agent that calls a hosted model all day, because the model is not running there. On a Doprax ProVM, the entry tier is 1 vCPU and 1 GB of memory, which is the shape this workload wants: small, always on, reaching out to an API when it needs to think.
Stop pricing the GPU. Price the VM, grab a free key, and ship the thing: https://www.doprax.com/products/virtual-machines
