What It Costs to Run an AI Agent: Real Monthly Numbers for Small Businesses

Running an open-source AI agent like Hermes costs between $5 and $75 a month for most small businesses: $0 for the software, $5-60 for model tokens depending on how much you use it and which model you pick, $2-5 for electricity if you host it on a Mac Mini, and whatever you optionally spend on search and browser tools. The number that surprises people is how low the floor is. A one-person company doing email triage and a daily report on a cheap model can stay under $10 a month, all in.

I've been running Hermes Agent on a Mac Mini since spring 2026, and I help small businesses set these systems up, so the numbers below come from actual invoices, not vendor math. Here's the full breakdown, line by line.

What is Hermes Agent, and what are you actually paying for?

Hermes Agent is an open-source AI agent framework from Nous Research, released under the MIT license. It runs on your own hardware - a Mac Mini, a $5 VPS, whatever you have - and connects to an LLM through an API key you bring yourself (BYOK, bring your own key). It reads your email, runs scheduled jobs on its built-in cron scheduler, does research, and talks to you over Telegram or Slack. The software itself costs nothing. That's not a freemium trick; the code is on GitHub and the license is genuinely permissive.

What you pay for is the thinking. Every time the agent reads an email or drafts a report, it sends text to a language model, and that model bills you per token. A token is roughly three-quarters of a word. That's the whole cost structure: compute you already own, plus model usage you rent by the sip. If you want the broader picture of what an agent like this actually does all day, I wrote up ten real Hermes use cases for small businesses earlier.

How much do LLM tokens cost for an AI agent?

Four model tiers are relevant for small business work in late 2026, and the spread between them is enormous.

• Claude Sonnet 5 (Anthropic) listed at $2 per million input tokens and $10 per million output during its introductory period through August 2026, with standard pricing reported at $3/$15 from September. Pricing guides disagree on whether the increase landed, so check the current Anthropic price sheet before budgeting - I calculate both below.
• Gemini 3.8 Flash (Google) is $0.75 per million input and $3.75 per million output.
• DeepSeek V4.1 Flash is $0.15/$0.60 per million off-peak and $0.30/$1.20 at peak hours, per DeepSeek's published pricing. Sub-dollar either way.
• A local model via Ollama is $0 per token. You pay only electricity, and you need serious RAM to make it good - more on that below.

Now the usage side. In plain terms, the workload I'm modeling is: 50-100 emails triaged a day, a handful of research questions, and two or three scheduled reports. That works out to roughly 0.2M input and 0.05M output tokens per day, or 6M input and 1.5M output over a month. The bill:

• Claude Sonnet 5: $27/month at the intro rate, $40 at standard pricing
• Gemini 3.8 Flash: about $10/month
• DeepSeek V4.1 Flash: $1.80 off-peak, $3.60 if everything runs at peak - and since Hermes's cron scheduler lets you run bulk jobs at 2am, off-peak is the natural setup

That's the honest range I quote clients: light daily use on a mid-tier model lands at $5-15 a month, an active operation with a heavier workload lands at $20-60, and only heavy agency-style usage on a frontier model pushes past $100. If someone tells you an agent needs a $500/month budget, they're either running a much bigger operation than yours or padding the quote.

One thing imo most guides get wrong: they price everything on frontier models. Triage and summarization don't need Claude. I run my own inbox triage on a Flash-class model and keep Sonnet for the tasks where judgment actually matters. Routing work to the cheapest model that handles it is the single biggest lever on this bill.

How much does the hardware and electricity cost?

Almost nothing, if you pick the machine everyone picks. The popular hosting choice for Hermes is a Mac Mini because Apple Silicon sips power. The base M4 Mini idles at 4-7 watts and draws 20-35 watts during active inference. At the US average electricity rate of 18.44 cents per kWh (EIA Electric Power Monthly, May 2026), a Mini mostly idling with a few hours of agent activity costs about $1-2 a month. Even at European rates you're under $5. A $5-10/month VPS works too if you don't want hardware in the office - the agent doesn't care where it lives.

The Mini itself runs $599 new. Amortized over three years that's about $17/month, but you were maybe buying a computer anyway. I mention it because "you need a dedicated server" is a common upsell and it's nonsense at this scale.

One hardware caveat: the $599 base Mini is perfect as a lightweight gateway to cloud models, but a true $0-token local operation needs memory. Capable local models through Ollama want 32-64GB of unified memory, which means a $1,600+ machine before the token savings mean anything. Local models are a privacy play, not a savings play, at SMB scale.

What do the extra tools cost?

Hermes's setup flow, hermes setup --portal, covers a model plus the Tool Gateway's four hosted services - web search, image generation, text-to-speech, and browser - under one OAuth. Note the Tool Gateway is only those four utilities; email reading comes from Hermes's mail integrations and scheduled jobs come from its built-in cron, not from the portal bundle. Beyond the portal:

• Web search: a Firecrawl or Tavily API key, typically free tier or $10-30/month if your agent researches daily
• Browser automation: usually unnecessary for a first deployment; the plain web fetch path handles most of it
• Honcho, the long-term memory layer (honcho.dev): free tier to start, or self-host if you're privacy-sensitive

None of these are required on day one. My advice: start with the model bill only, then add tools when a specific job demands them. If you're setting up for the first time, my skills-to-install-first guide covers which ones actually earn their keep.

How does this compare to the SaaS you're already paying for?

Here's the comparison that actually convinced me to move. Zylo's 2025 SaaS Management Index puts average SaaS spend at $4,830 per employee per year - about $402 per employee per month - and companies under 20 employees spend the most per head, around $8,000 a year. VendorBenchmark's 2026 dataset puts small business stacks at 60-120 applications. More than half of those licenses go unused for 90+ days, per the same Zylo data.

An agent doesn't replace all of that. Your accounting package survives. But the pattern I keep seeing in client audits is three or four subscriptions - a monitoring tool, a scheduling assistant, a report generator, a Zapier plan - that one Hermes instance with $20/month of tokens absorbs entirely. Four subscriptions at $40 each is $160/month. The agent doing the same work costs $25/month including electricity. Put next to Zylo's $402/employee/month, a single agent replacing even two seats of niche tooling pays for itself several times over. And it costs tokens, not seats, which matters if you intend to grow.

The honest accounting includes your time. Setup is a weekend or two if you're technical, or a consultant invoice if you're not. The agent still needs supervising - it removes the typing, not the decisions. Anyone selling you "an employee for $20/month" is lying; what you're buying is leverage on work you'd otherwise do manually.

Where do AI agent budgets leak?

Three failure modes I've seen, twice each, in real deployments:

• The agent gets stuck. A scheduled job errors, the agent retries, errors again, retries again - each attempt burns tokens, and left alone it can loop all night. Twenty wasted retries a day carrying a fat 10k-token context is 6M extra input tokens a month, which is $18/month on Sonnet standard pricing and real money on nothing. Cap retries and watch the first week of usage graphs.
• Bloat in the prompt. If you let the agent carry an entire email history into every triage call, you pay for those tokens on every single message. Trim what it reads.
• Frontier model by default. A client ran their first month on a top-tier model out of the box - a shared inbox summary plus one weekly report, nothing exotic - and got a $94 invoice. Same workload on a Flash-class model: $11. They nearly cancelled the whole project over a configuration choice.

Set a hard billing limit in your provider's dashboard on day one - Anthropic, Google, and OpenRouter all support spend caps. For zero runaway-bill risk, route fallback work to a local Ollama model where the marginal token cost is literally nothing.

What's the total monthly cost, then?

For a typical small business: $0 software, $5-60/month in tokens depending on workload and model choice, under $5 in electricity or a $10 VPS, and $0-30 in optional tool keys. On a mid-tier stack - Gemini Flash on a Mac Mini - budget $30/month and you'll rarely be surprised. If you insist on a frontier model for everything, expect $50-60. Either way, compare that to the $200-1,000/month SaaS stack doing less work, and the economics explain themselves. And if you're handing credentials to any agent, read my agent security guide before you connect anything to your bank.