LLM and agent cost planning
Running a model is more than a token price. Tell us what you want to run, where it should run, and how you plan to use it. We’ll show the model, host, runner, hardware, interface, and usage assumptions that drive the bill before you build.
Not for you if you need a live gateway, request tracing, or exact billing guarantees. AIStackPicker is a pre-build planner.
Example monthly estimate
| Host: Contabo Cloud VPS S | $6.25 |
| Model: Gemini 3 Flash | $19.00 |
| Runner: OpenClaw | — |
| Interface: Messaging only | — |
| Messaging: Telegram | — |
| Capabilities: Web search, Memory | — |
| Extras: Tailscale | — |
| Estimated total | $25.25/mo |
Tell us whether you are running chat, a local model, or an always-on agent. We’ll shape the estimate around that workload, because each setup needs different parts.
Most AI cost calculators stop at tokens. We price the rest of the stack too: compute, hardware, runners, interfaces, capabilities, and usage assumptions.
The same model can be cheap through an API, expensive on an idle rented GPU, or practical on hardware you already own. The setup matters.
A chat app usually answers one request. An agent may plan, call tools, read files, browse, retry, write results, and check its own work before it stops.
We separate those costs because one request can turn into many model calls, a growing context window, tool output, logs, memory, and always-on hosting.
For chat, count tokens. For agents, count steps.
Tell us whether you want an API, a rented GPU, local hardware, or a hybrid setup. We’ll keep the cost models separate so you are not comparing tokens against idle GPU hours or hardware you already own.
A rented GPU charges by time, not tokens. Monthly equivalent means hourly price × 730 hours. If the GPU sits idle, it can still cost money. GPU hosting vs API →
Local inference has no token bill, but it is not free. VRAM, unified memory, quantization, context length, power, and setup all matter. Check local hardware fit →
A few builds priced with AIStackPicker, showing how a full stack adds up. Open one to see every part and its monthly estimate.
| Host | Contabo Cloud VPS S |
| Model | Claude Sonnet 4.6 |
| Runner | OpenClaw |
| Interface | Messaging only |
| Messaging | Telegram |
| Capabilities | Web search, Memory |
| Extras | Tailscale |
| Host | Lambda A100 SXM |
| Model | Gemma 3 27B |
| Runner | vLLM |
| Interface | Command line |
| Host | Apple Mac mini (M4) · 32 GB |
| Model | Gemma 3 27B |
| Runner | Ollama |
| Interface | Command line |
These guides explain the decisions behind the estimate: when local makes sense, when a rented GPU gets expensive, and why agents can cost more than simple chat.
AI prices move. Cloud GPU availability changes. Local hardware behaves differently by model and runner. Agents are hard to predict because they may take more steps than expected.
We do not hide that uncertainty behind one confident number. We show what was priced, what was assumed, and where the estimate can break.
Tell us whether you are running chat or an agent. Then choose whether it belongs in the cloud, on local hardware, or somewhere in between. We’ll price the parts that change the bill.