How Much Does It Cost to Run an AI Agent per Month?

The Honest Answer Is $6 to $80, and the Spread Is Almost Entirely Model Spend

A personal AI agent running 24/7 on a small VPS with an OpenRouter or Anthropic key typically lands somewhere between $6/month at the low end and $80/month at heavy usage. The infrastructure is nearly a rounding error on that bill: it is $4 to $8 for a small VPS every month regardless of how much you talk to the agent. Everything above that floor is model spend, which scales with how many tokens the model reads and writes on your behalf.

If you have been trying to reverse-engineer that number from calculator pages that quote $200 to $10,000/month, the confusion is real. Those numbers describe enterprise call-center agents handling tens of thousands of conversations a month, not one person's assistant on Telegram. This post is the breakdown for the second case: one owner-operator, one messaging channel, the honest math.

Why the Bill Splits Into Two Halves

The monthly cost of an AI agent is the sum of two independent line items:

Infrastructure: a small Linux server that stays online, holds conversation state, and runs the messaging gateway. Fixed cost.
Model spend: pay-per-token calls to the model provider (OpenRouter, Anthropic, OpenAI). Variable cost, scales with usage.

You almost never run the model itself on your own machine for a personal agent. Local LLM inference needs a GPU or 16-32 GB of RAM, which pushes the hardware bill above what pay-per-token would cost you at realistic personal usage. The rest of this article assumes the model lives at a provider and your VPS just talks to it.

That split is what makes the total honest to estimate. The infrastructure floor is knowable in advance; the model spend is a function of your message count and token length.

Line 1: The Infrastructure Floor

An API-driven agent on Telegram, Signal, or Slack needs surprisingly little hardware. The minimum baseline is one to two vCPUs, 1 to 2 GB of RAM, and 20 to 40 GB of SSD. Most runtimes, including Hermes Agent, will run comfortably inside 2 GB of RAM.

Real 2026 monthly numbers for that spec:

Hetzner CX22 (2 vCPU, 4 GB RAM, 40 GB NVMe): about €3.79/month in the EU, roughly $4.59 to $4.99 in the US regions. Includes 20 TB of egress. This is the default recommendation in self-hosting communities.
DigitalOcean Basic Droplet (1 vCPU, 1 GB RAM): $6/month, or about $12/month for the 2 GB tier. Predictable pricing and cleaner docs than Hetzner, at 2-4x the price for the same spec.
Oracle Cloud Always Free ARM: $0/month if your account survives the reclaim policy. Real free tier, real friction on account approval.
Raspberry Pi 5 at home: $80 one-time plus about $2 to $4/month in electricity for 24/7 uptime. Cheaper long-term, worse for uptime if your home internet is not stable.

For a first estimate, use $6/month as your infrastructure floor. Anything from a Hetzner CX22 to a small DigitalOcean droplet lands there or below. If you want the trade-offs in detail, we compared them in Cheap VPS for AI Agents.

Line 2: The Model Spend

This is where the range widens. Model providers charge per million tokens of input (what the model reads, including the growing conversation history) and per million tokens of output (what it writes back). A token is roughly 4 characters of English, or about 0.75 of a word.

The rates that matter for a personal agent in mid-2026:

Claude Haiku 4.5: $1.00 per million input tokens, $5.00 per million output tokens. The cheap workhorse for most day-to-day chat.
GPT-4o: about $2.50 per million input tokens on OpenRouter, $10 per million output. Sits in the same performance band as Claude Sonnet.
Claude Sonnet 4.6: $3.00 per million input, $15 per million output. Reach for it when the task actually needs the extra reasoning.
Gemini Flash: from $0.075 per million input tokens on OpenRouter. Cheapest of the mainstream options, useful as a fallback tier.

Two things blow past what beginners expect:

Every message replays the whole conversation. The model has no memory of its own. Each new message is processed as if all prior turns were prepended to the input, so by turn 10 a single reply is being priced against about 7x the tokens of turn 1 for the same output. This is why a chatty agent that never trims context can look expensive even when each individual message is short.
Output tokens cost 4-5x more than input. A verbose model that writes long essays for a one-line question is spending five times what a curt one does for the same reply quality.

What That Looks Like in Real Monthly Bills

Assume 4 characters per token, 750 tokens per typical short assistant reply, and system-prompt overhead in the low hundreds of tokens. Rounded, honest ballparks for three usage profiles on Claude Haiku 4.5 or a similar sub-dollar-input model:

Light usage (about 20 conversations/day, 6 turns each): roughly 3 to 6 million tokens/month. $5 to $12/month in model spend.
Medium usage (about 80 conversations/day, 8 turns each): roughly 15 to 25 million tokens/month. $25 to $50/month.
Heavy usage (about 200 conversations/day, 10 turns each, with voice transcription and TTS): 40 to 80 million tokens/month plus voice audio. $60 to $150/month.

Switch to Claude Sonnet 4.6 for everything and those numbers roughly triple. Switch to Gemini Flash and they roughly halve. Model choice is by far the biggest lever you have on the total.

Optional Line: Voice

If the agent takes voice messages and replies in voice, the bill grows by two more variable lines: STT (speech-to-text) to transcribe the incoming audio and TTS (text-to-speech) to synthesize the reply. Numbers to plan against:

OpenAI Whisper API: $0.006 per minute of audio in. A minute-a-day habit is under $0.20/month; a 30-minute-a-day habit is about $5/month.
ElevenLabs: usage-based, starts at $5/month for their Starter plan, real production use lands $22 to $99/month depending on characters synthesized. See ElevenLabs Voice Setup for Hermes Agent for the trade-offs.
Self-hosted alternatives: Whisper.cpp, faster-whisper, Piper, or Edge TTS. Free at the software level, they add a few GB of RAM to your VPS spec if you go that route.

For most owner-operators, voice adds $10 to $30/month to a text-only setup at typical usage.

Putting It Together: The Three Profiles

Rolling up everything above, the honest monthly cost of running a personal AI agent 24/7 in 2026 looks like this:

Light (text-only, cheap model, small VPS): $10 to $20/month. Infrastructure $6, model spend $5 to $12, voice $0.
Medium (mixed models, occasional voice, small VPS): $40 to $80/month. Infrastructure $6, model spend $25 to $50, voice $10 to $25.
Heavy (Sonnet-class model, daily voice, headroom RAM): $100 to $200/month. Infrastructure $12, model spend $60 to $150, voice $20 to $40.

The bill is dominated by model choice, not hosting. A user who moves from Sonnet-for-everything to Haiku-for-chat-and-Sonnet-for-hard-questions can cut their monthly total by half or more without losing the reasoning they actually need.

The Hidden Line: Your Time

The self-hosted numbers above assume the agent stays up on its own. In practice the first month usually costs an evening of setup, and every OS or runtime upgrade costs another hour. Rate limits, expired API keys, or a Docker container that quietly restarts each night (a common problem covered in Hermes Agent Docker Container Keeps Restarting) each cost their own diagnosis time.

The way to price that fairly is per hour of your own labor. If babysitting a self-hosted agent costs you two hours a month and your time is worth $50/hour, that is $100/month you are paying yourself in kind. It rarely shows up on a spreadsheet, but it is the reason most owner-operators eventually move to a managed option.

The Managed Alternative

If you would rather not run the VPS, keep the OpenRouter account, or track any of this: Hermify hosts a managed Hermes Agent on Telegram with persistent memory that stays yours. You bring your own model key (BYOK) on the entry tier, so the model spend still shows on your bill and you still control it, but the infrastructure line disappears entirely. For readers researching cost specifically against a brand-aware equivalent, we broke down our own tiers in How Much Does Hermes Agent Cost.

That trade is the honest conclusion of any monthly-cost article: the raw infrastructure floor of a self-hosted agent is cheap, the model spend is the same either way, and the hidden variable is how much of your own attention the setup consumes. Do the math on all three lines before you pick.