CostsCosts
Model or machine?
An agent's monthly bill has two parts: the computer it lives on, and the language model it calls. Most comparisons price only the computer. This page puts both side by side for every ready-made workload, so you can see which choice moves your bill.
Short answer: For a personal agent that stays on, the machine costs $5.35 – $250.00 a month depending on the provider, and the model $7.30 – $488.07 depending on the model. For every workload here that calls a model, a frontier model costs at least 11 times the cheapest machine that runs it.
Every workload, both halves
A month at list prices. Machine: cheapest to dearest provider that runs it as asked. Model: median price of hosts that keep no data.
| Workload | Machine | Tokens a month | Cheap models | Mid-priced | Frontier | Frontier ÷ cheapest machine |
|---|---|---|---|---|---|---|
| A personal agent that stays on (Hermes, OpenClaw) | $5.35 – $250.00 | 144M | $7.30 – $7.46 | $61.26 – $160.83 | $204.19 – $488.07 | 38× – 91× |
| A coding agent for one developer | $6.60 – $82.42 | 360M | $18.25 – $18.65 | $153.14 – $402.08 | $510.48 – $1,220.18 | 77× – 185× |
| A code interpreter inside your product | $2.42 – $756.00 | No model | ||||
| A browser agent for a small team | $36.30 – $1,159.80 | No model | ||||
| Self-hosted n8n for a small team | $5.35 – $250.00 | 45M | $2.28 – $2.33 | $19.14 – $50.26 | $63.81 – $152.52 | 12× – 29× |
| A Discord or Slack bot for one server | $4.08 – $250.00 | 90M | $4.56 – $4.66 | $38.29 – $100.52 | $127.62 – $305.05 | 31× – 75× |
| Claude Code for a team of five | $28.80 – $412.09 | 1,800M | $91.26 – $93.24 | $765.72 – $2,010.42 | $2,552.40 – $6,100.92 | 89× – 212× |
| A scraper that runs four times a day | $0.00 – $20.00 | No model | ||||
| Your own remote MCP server | $4.08 – $250.00 | No model | ||||
| A: 1,000 short tasks | $1.21 – $50.00 | No model | ||||
| B: one always-on box | $6.60 – $250.00 | No model | ||||
| C: three hours a day | $2.88 – $164.90 | No model | ||||
What each row says
- A personal agent that stays on (Hermes, OpenClaw). Machine $5.35 – $250.00 (OVHcloud VPS to Tensorlake, 22 providers); model $7.30 – $488.07 (MiniMax M3.1 Flash to GPT-6 Astra). Even the cheapest model costs more than the cheapest machine; a frontier model costs 38× to 91× as much. The model is the decision. Price it yourself
- A coding agent for one developer. Machine $6.60 – $82.42 (Contabo to Amazon Lightsail, 23 providers); model $18.25 – $1,220.18 (MiniMax M3.1 Flash to GPT-6 Astra). Even the cheapest model costs more than the cheapest machine; a frontier model costs 77× to 185× as much. The model is the decision. Price it yourself
- A code interpreter inside your product. Machine $2.42 – $756.00 (Agent 37 to Hetzner Cloud, 27 providers). No model in this workload, so the machine is the whole bill: $2.42 – $756.00 depending on the provider. Price it yourself
- A browser agent for a small team. Machine $36.30 – $1,159.80 (Agent 37 to Modal, 27 providers). No model in this workload, so the machine is the whole bill: $36.30 – $1,159.80 depending on the provider. Price it yourself
- Self-hosted n8n for a small team. Machine $5.35 – $250.00 (OVHcloud VPS to Tensorlake, 22 providers); model $2.28 – $152.52 (MiniMax M3.1 Flash to GPT-6 Astra). On a cheap model the machine is still the bigger half. On a frontier model, the model costs 12× to 29× as much as the cheapest machine: the model is the decision. Price it yourself
- A Discord or Slack bot for one server. Machine $4.08 – $250.00 (HostUp to Tensorlake, 23 providers); model $4.56 – $305.05 (MiniMax M3.1 Flash to GPT-6 Astra). Even the cheapest model costs more than the cheapest machine; a frontier model costs 31× to 75× as much. The model is the decision. Price it yourself
- Claude Code for a team of five. Machine $28.80 – $412.09 (Boat to Amazon Lightsail, 23 providers); model $91.26 – $6,100.92 (MiniMax M3.1 Flash to GPT-6 Astra). Even the cheapest model costs more than the cheapest machine; a frontier model costs 89× to 212× as much. The model is the decision. Price it yourself
- A scraper that runs four times a day. Machine $0.00 – $20.00 (Modal to Vercel Sandbox, 27 providers). No model in this workload, so the machine is the whole bill: $0.00 – $20.00 depending on the provider. Price it yourself
- Your own remote MCP server. Machine $4.08 – $250.00 (HostUp to Tensorlake, 23 providers). No model in this workload, so the machine is the whole bill: $4.08 – $250.00 depending on the provider. Price it yourself
- A: 1,000 short tasks. Machine $1.21 – $50.00 (Agent 37 to Freestyle, 27 providers). No model in this workload, so the machine is the whole bill: $1.21 – $50.00 depending on the provider. Price it yourself
- B: one always-on box. Machine $6.60 – $250.00 (Contabo to Tensorlake, 22 providers). No model in this workload, so the machine is the whole bill: $6.60 – $250.00 depending on the provider. Price it yourself
- C: three hours a day. Machine $2.88 – $164.90 (Agent 37 to E2B, 24 providers). No model in this workload, so the machine is the whole bill: $2.88 – $164.90 depending on the provider. Price it yourself
The assumptions
Token volumes are our assumptions, not measurements. Each is messages a day × thousands of tokens per message × the days a month the machine works. An agent re-sends its instructions and history on every model call, and one message can take several calls.
- A personal agent that stays on (Hermes, OpenClaw): 30 × 160k × 30 days = 144 million tokens.
- A coding agent for one developer: 20 × 900k × 20 days = 360 million tokens.
- Self-hosted n8n for a small team: 50 × 30k × 30 days = 45 million tokens.
- A Discord or Slack bot for one server: 100 × 30k × 30 days = 90 million tokens.
- Claude Code for a team of five: 100 × 900k × 20 days = 1,800 million tokens.
The mix is 5% output and 95% of input read from cache (4.75% new input, 90.25% cached input). Model prices are list prices from TokenWatch, read 2026-10-31, at the median of the hosts that keep no data. Machine prices are each provider's list prices, checked 29 October 2026, counting only providers that run the workload as asked. Taxes, egress and anything a provider does not publish are left out. A plan you already pay for can replace the model line; the calculator shows that case. How the model line works.