How Much Does Hermes Agent Cost?

The short answer

Hermes Agent costs $0 for the software and somewhere between about $3 and $60+ per month to actually run, depending on three things: where it runs, which AI model it calls, and how much it works. For most people the model bill, not the hosting bill, is the part that moves. Model prices below were checked in October 2026.

The four parts of the bill

1. Software: $0. The Hermes Agent repository is MIT licensed. There is no license fee. See Is Hermes Agent free? for the full answer.
2. Hosting: the machine that keeps the agent online. $0 on your own PC, a few dollars on a VPS or a managed plan.
3. Model usage: tokens billed by whichever provider you connect. This is the variable cost.
4. Your time: setup, updates, restarts. Zero for managed, real hours for DIY.

Hosting options and what they cost

Your own computer. No rent, but the machine has to stay on for the agent to answer messages or run scheduled jobs, so you pay in electricity and in laptop sleep settings. See the system requirements before you commit a machine.

A VPS. Budget servers typically run about $4–$7 per month for a small instance, plus the time to install, secure and update Hermes yourself. Our comparison is in Cheapest Hermes Agent Hosting and the server picks are in Best VPS for Hermes Agent.

Managed on OpenClaw Launch. There is a free 30-minute trial with no credit card. Lite is $3 for the first month, then $6 per month (or $60 per year, $5 per month effective): 1 vCPU, 2 GB RAM, 10 GB storage, 1 instance, and $1 per month of AI credits. Pro is $20 per month (or $200 per year): 2 vCPU, 4 GB RAM, 40 GB storage, up to 3 instances, and $10 per month of AI credits. The credits are spent through OpenRouter at standard rates with no markup. You can also bring your own OpenRouter key or connect a ChatGPT subscription. Details are on the pricing page.

Nous Portal. Nous Research sells its own subscription that bundles models, a tool gateway and Hermes Cloud hosting, with a free tier limited to free models. Plans and prices change, so check Nous's own site and read our Nous Portal guide for how it compares.

Why agents burn more tokens than a chat

In a chat window you send a question and get an answer: one model request. An agent works in a loop. It reads your message, calls a tool (search the web, read a file, run a command), reads the result, and decides what to do next. Each step is a separate model request, and each request resends the system prompt, the conversation so far and every earlier tool result. A single web page or file read can add thousands of tokens that are then carried into all following steps.

So the honest unit of cost is not “per message” but “per task”, and it varies widely. Prompt caching, offered by some providers, can reduce the price of the repeated part. The estimates below ignore caching, so treat them as a conservative ceiling for a task of this shape.

Model prices (checked October 2026)

Per-million-token prices as listed on OpenRouter (read from its public models API on October 7, 2026). Providers can change them at any time.

The example task is an assumption, not a measurement: one user message that triggers 8 model requests, each with about 12,000 input tokens (history plus tool output), and 4,000 output tokens in total. That is 8 × 12,000 = 96,000 input tokens and 4,000 output tokens. Cost = 0.096 × input price + 0.004 × output price. For GPT-5.6 Sol: 0.096 × $2.00 + 0.004 × $10.00 = $0.192 + $0.040 = $0.232. For GLM-5.3 Flash: 0.096 × $0.15 + 0.004 × $0.50 = $0.0144 + $0.002 = $0.016. Your real tasks may be several times smaller or larger. Free models exist on OpenRouter (16 were listed on the date above), but they are rate-limited and weaker, which suits testing more than production. Which models we feature is on the models page.

The ChatGPT subscription route

If you already pay for ChatGPT Plus, Pro or Team, OpenClaw Launch lets you connect that subscription and run GPT-5.6 with no extra API charge, so the model line of your bill can be $0 beyond what you already pay OpenAI. Usage limits come from your ChatGPT plan. Setup is in Hermes Agent with a ChatGPT subscription.

Monthly cost by setup

Three worked examples

All use the example task above ($0.016 on GLM-5.3 Flash, $0.232 on GPT-5.6 Sol) and 30 days per month. The totals are usage estimates: they count tokens at OpenRouter rates and leave out payment fees. On OpenClaw Launch, usage beyond the included credits is bought as a $5, $10 or $20 credit top-up, which costs $5.60, $10.80 or $21.20 with the card-processing fee, so your actual payments come in those blocks. With your own OpenRouter key you pay OpenRouter directly instead.

1. Light user: 5 tasks a day on a cheap model

150 tasks × $0.0164 (the unrounded GLM-5.3 Flash task cost) = $2.46 of tokens. Managed Lite costs $6 and includes $1 of credits, leaving $1.46 to pay. Total about $7.50 per month ($4.50 in the first month at the $3 intro price). Staying within free models would bring this to the hosting cost alone.

2. Daily user: 20 tasks a day, mostly cheap, some premium

600 tasks a month. 540 on GLM-5.3 Flash: 540 × $0.0164 = $8.86. 60 on GPT-5.6 Sol: 60 × $0.232 = $13.92. Tokens total $22.78. On Lite: $6 + ($22.78 − $1 credits) = about $27.80. On Pro: $20 + ($22.78 − $10 credits) = about $32.80, which buys extra RAM, storage and instances rather than cheaper tokens. Connecting a ChatGPT subscription for the premium share would remove the $13.92 line.

3. Heavy automation: 100 scheduled tasks a day

3,000 tasks a month, all on GLM-5.3 Flash: 3,000 × $0.0164 = $49.20. On Pro: $20 + ($49.20 − $10) = about $59.20. The same load on GPT-5.6 Sol would be 3,000 × $0.232 = $696, which is why heavy automation should run on a cheap model or a subscription you already hold.

How to cut the bill

• Use a cheap model for routine turns and switch to a premium one only for hard tasks. In the table, the gap is roughly 14 times between GLM-5.3 Flash and GPT-5.6 Sol per task.
• Cap context. Start fresh sessions for new topics; long histories are resent on every request.
• Use free models while testing prompts and skills, then move production to a paid model.
• Watch scheduled jobs. A cron task that runs every few minutes multiplies cost faster than any chat.
• Use a subscription you already pay for if you have ChatGPT Plus, Pro or Team.
• Compare providers in Hermes Agent with OpenRouter and best models for Hermes Agent.

Does OpenClaw cost the same?

Structurally, yes. OpenClaw is MIT licensed and its repository states there is no paid tier from the project, so, as with Hermes, you pay only for hosting and model tokens. On OpenClaw Launch the two frameworks are priced identically. Model costs follow the model you choose, not the framework. For OpenClaw-specific hosting prices see Cheapest OpenClaw Hosting.

Frequently Asked Questions

Related Guides

• Cheapest Hermes Agent Hosting — hosting-only price comparison
• Best VPS for Hermes Agent — server options
• Is Hermes Agent Free? — the licensing answer
• Nous Portal — Nous Research's subscription
• Hermes with a ChatGPT subscription — no extra API cost
• Hermes Agent System Requirements — what hardware you need
• Managed vs Self-Hosted — the trade-offs
• Cheapest OpenClaw Hosting — the OpenClaw side