FAQ
Plain answers about what it costs to run an LLM or AI agent, where it can run, why tokens and GPU time are different meters, what parts must fit together, and how the prices on this site are sourced.
The basics
An AI agent stack is the set of parts that let an LLM or agent run. The usual parts are a host, an AI model, a runner, an interface, capabilities, and optional extras. The host is where it runs. The model is what reasons or writes. The runner is the software that runs the model or agent.
An LLM is a model you use directly: chat, an app, a CLI, or an API. An agent is a model plus software that can use tools, memory, schedules, or messaging, and usually runs on its own. Choose an LLM package for direct use, an Agent package for something that acts on its own. Either can run in the cloud or on your own hardware.
No. AIStackPicker is for AI assistant and agent stacks generally. OpenClaw and similar tools are examples of why people need clearer cost and compatibility comparisons. The catalog supports whatever runners, models, hosts, and messaging channels are in the data.
Cost
The monthly cost usually has two parts: fixed cost and model usage cost. Fixed cost includes the host and required services. Model usage cost depends on how much the assistant reads and writes. AIStackPicker estimates both and shows a typical range.
The cheapest always-on setup is the lowest-cost compatible stack that can stay online and meet the job's requirements. A very cheap host may fail if the runner needs more memory. A cheap model may not stay cheap under heavy use.
Hosting keeps the assistant online. Model usage pays for the AI model to read, reason, and write. If the assistant works often, model usage can cost more than the host.
No. The cheapest setup is useful only if it works for the job. AIStackPicker tries to find the cheapest reliable compatible setup, then lets you trade cost, convenience, and control in the Builder.
Usually not. Processing tokens always costs compute, and you pay for it one of three ways: per token to an API provider, as GPU rental when you self-host in the cloud, or as electricity when you self-host on hardware you already own. Renting a GPU is not free tokens. The rental is the token cost, and run 24/7 it often costs more than the API. The only genuine cut is hardware you already have, and only if it fits the model and stays on.
Tokens, GPU time, and hardware
A model API charges by tokens. You pay for the input and output the model processes, and an idle month costs almost nothing. A rented GPU charges by time. You pay for every hour it runs, whether it answers one request or a thousand. Local hardware shifts the cost again: no token bill, but you pay in the machine, power, and setup. These are three different meters, not three prices for the same thing.
Monthly equivalent means the hourly price times 730 hours, the number of hours in an average month. That figure assumes the GPU runs all month. If you stop it between jobs, the actual bill is lower. It is also why a rented GPU can look expensive next to an API for light use: you are paying for idle hours too.
A chat setup usually answers one request. An agent may take many steps before it stops: planning, calling tools, reading files, retrying, and checking its own work. Each step can use the model again, and always-on agents also need a host, logs, and budget limits. Agents are not priced like chat.
Usually not. Many agents run their logic on a small, cheap host and call an API model only when they need to think. A GPU is only needed when the model itself runs on your own hardware. Do not rent a GPU only because the workload is called an agent.
Hosting and self-hosting
Yes, if your computer supports the runner and is on when the assistant needs to work. Local setups can be useful for privacy and control. They are not always the right choice for scheduled tasks, monitoring, or anything that must work when your computer is off.
You need a VPS or managed host if the assistant must stay reachable when your computer is off. A VPS is a rented server. It can be cheaper than managed hosting, but it usually requires more setup and maintenance.
Self-hosted means you run the assistant on your computer or rented server. Managed means a provider handles more of the setup or operation. Self-hosted usually gives more control. Managed usually reduces setup work.
Capabilities and models
Browser automation uses more memory and creates more security risk than simple chat or scheduled summaries. A setup that works for light tasks may not work for web research or form-filling. AIStackPicker treats browser automation as a real capability with real requirements.
No. Benchmarks may help compare models, but the product is broader than benchmark scores. A useful assistant stack also depends on cost, compatibility, host requirements, runner support, interfaces, and how current the data is.
Data, sources, and trust
AI pricing and hosting plans change. A price without a date can be misleading. AIStackPicker shows where a price came from and when it was last checked so you can judge the estimate.
Stale data is too old to trust as a default recommendation. AIStackPicker may still show stale facts for context, but it warns before using them in an estimate or recommendation.
No. Affiliate status does not affect ranking, compatibility, price estimates, warnings, or recommendations. Some outbound provider links may earn AIStackPicker a commission, but the picker logic does not read payout data.
For agents and developers
Yes. The MCP/API surface lets agents query the same data used by the human catalog and Builder. The endpoint is read-only and does not return affiliate URLs.