What does it cost to run an AI agent 24/7?
At API list prices, an AI agent running 24/7 costs about $6 a month on an efficient model such as GPT-6 Luna and $189 on Claude Opus 5.5 when its days look like our median agent-day. An agent as busy as our busiest tenth of days, every day, costs $71 to $2,095: ten to twelve times as much.
Cost per agent per month
| Agent | Opus 5.5 | Sonnet 5.5 | DeepSeek V4.1 Flash | GPT-6 Luna |
|---|---|---|---|---|
| Typical: our median day × 30 | $189 | $95 | $7 | $6 |
| Busy: our 90th-percentile day × 30 | $2,095 | $1,048 | $74 | $71 |
Real requests from September 13 to October 12, 2026, priced at each model's list price with the cache hits they actually had. One API key on one UTC day is one agent-day. Prices are in the method section below.
What drives the bill
Every step re-sends the conversation, tool results included, so the 100th request in a session carries five times the context of the first.
| Driver | What we measured |
|---|---|
| Context growth | Requests over 100K input tokens were 32% of requests and 55% of the cost at Sonnet 5.5 prices. |
| Heartbeats and polling | A lone wake-up found 29% of its 14,200 input tokens in the cache, against 87% overall, because the cache had expired. At Sonnet 5.5 prices that is $0.033 a wake-up, twice a median request, or $47 a month at one every 30 minutes. |
| Tool calls | Output was 1.1% of tokens but 23% of the Sonnet 5.5 cost, and each tool result returns as input on the next step. |
| Retries | A retried step is billed again with its whole context. Our figures count completed requests only. |
Five settings that cut the cost
Without cache hits, the requests in our data would have cost 4.5 times as much at Sonnet 5.5 prices. Anthropic's cache lasts 5 minutes, or an hour at twice the write price, which pays off for agents that wake every 5 to 60 minutes.
| Setting | OpenClaw | Hermes Agent |
|---|---|---|
| Cheaper model for side work | heartbeat.model, subagents.model and compaction.model under agents.defaults; subagents otherwise inherit the main model | auxiliary.<task>.model for compression, vision and other side tasks, which use the main model by default; delegation.model; cron.model |
| Context compaction | On by default near the context window, keeping the last 20,000 tokens | compression.threshold, 0.50 of the window by default |
| Heartbeat interval | heartbeat.every, 30m by default, "0m" turns it off; activeHours; isolatedSession cuts a run from about 100K to 2-5K tokens | None by default; /heartbeat adds one; cron jobs with --no-agent call no model |
| Cache lifetime | cacheRetention "long" for the 1-hour cache | prompt_caching.cache_ttl "1h" |
| Spend cap | None; /usage cost reports spend | None; /usage and /insights report a lower-bound estimate |
From the OpenClaw docs (models, compaction, heartbeat, prompt caching) and the Hermes Agent configuration docs, checked October 13, 2026. Neither agent caps dollars, so the cap belongs on each agent's API key.
Budgeting a fleet of agents
Budget per agent, not per fleet. Give each agent its own API key with its own monthly cap, run it for a week, and set the cap from its spend per active day times 30. Agents with spiky work fit a prepaid balance with a monthly limit better than a fixed plan.
| Fleet | Opus 5.5 | Sonnet 5.5 | DeepSeek V4.1 Flash | GPT-6 Luna |
|---|---|---|---|---|
| 5 agents: 4 typical, 1 busy | $2,850 | $1,430 | $100 | $90 |
| 10 agents: 8 typical, 2 busy | $5,700 | $2,850 | $200 | $190 |
| 25 agents: 20 typical, 5 busy | $14,300 | $7,130 | $510 | $470 |
About, per month at list prices, every agent on one model. One busy agent in five is a planning assumption: count the agents that work all day and adjust.
How we measured
We priced every completed request on the Standard Compute gateway from September 13 to October 12, 2026. The median agent-day had 101 requests and 4.3M tokens; the 90th-percentile day had 775 requests and 79M tokens over a median of 14 active hours. A month is 30 days, since an always-on agent works weekends.
Agents don't name themselves reliably in request logs, so the figures pool all agent traffic, not OpenClaw or Hermes alone, and keys active 20 or more hours a day were too few to publish as a group. The gateway runs around the clock: no UTC hour carried less than 3.3% or more than 5.2% of requests.
List prices per million tokens on October 13, 2026: Opus 5.5 $4 input, $0.20 cache read, $20 output; Sonnet 5.5 $2, $0.10, $10 (Anthropic); DeepSeek V4.1 Flash $0.30, $0.006, $1.20 in peak hours and half that otherwise (DeepSeek); GPT-6 Luna $0.10, $0.01, $0.50, more above 272K input (OpenAI).
The catch
A key at its budget returns HTTP 402 until the plan renews or an admin raises it, so the agent stops: a scheduled job does not run and a Slack or Telegram bot goes quiet. Give the agents that must not stop a prepaid balance with automatic refill and a limit well above their normal month, and make every agent report a 402 somewhere a person will see it. Our figures pool all agent traffic on our gateway, so run your own agent for a week before you size its key.
Where Standard Compute fits
Standard Compute gives each agent its own application key with its own plan, from Starter at $19 to Fleet at $4,999 a month, which stops at its budget while the other agents keep running. Spiky agents can use Pay as you go instead: a prepaid company balance at usage plus 15%, with a monthly limit per key. OpenClaw and Hermes connect as an OpenAI-compatible provider at https://api.stdcmpt.com/v1. Creating the company is free, and billing starts when the first key's plan is active.
FAQ
How much does it cost to run OpenClaw 24/7?
OpenClaw is free; the model calls are the cost. At list prices an agent with typical days costs from about $6 a month on GPT-6 Luna to $189 on Claude Opus 5.5, and one busy all day costs ten to twelve times that. The default heartbeat every 30 minutes adds about $47 a month on Sonnet 5.5 unless you lengthen it or set agents.defaults.heartbeat.every to "0m".
What does Hermes Agent cost per month to run?
Hermes Agent is free and MIT-licensed, so the bill is the models it calls. Running all day on its messaging gateway or cron jobs, expect the same ranges as any always-on agent: single dollars a month on an efficient model for typical use, hundreds to a couple of thousand dollars on a frontier model when it works through every day.
Why does an idle agent still cost money?
Every wake-up re-sends the agent's context: system prompt, tools and recent history. Wake-ups spaced further apart than the provider's cache lifetime, five minutes by default on Anthropic, pay close to full input price each time. In our data a lone wake-up cost twice a median request.
Is a cheap model good enough for an always-on agent?
For heartbeats, triage, summaries and routine tool calls, usually yes, at a tenth or less of a frontier model's cost. Keep a frontier model for the steps that need judgment, and judge by cost per finished task: a cheap model that loops and retries can cost more than a stronger one that finishes.
Should each AI agent have its own API key?
Yes. A key per agent shows which agent spends what, lets you cap each one separately, and stops only that agent when it reaches its cap. On one shared key, a single runaway agent spends the whole budget and stops every other agent with it.
What happens when an agent reaches its budget on Standard Compute?
Its key returns HTTP 402 until the plan renews or an admin upgrades it, and every other agent keeps running. Plans have no overage charges. A key on Pay as you go runs from the prepaid company balance until it reaches its monthly limit or the balance runs out.