On 20 October 2026 the first OpenAI dot is included with ChatGPT Pro ($100, $200 or $500 a month) or Business Premium; Meta Muse has a free tier, $20 and $100; Gemini Spark needs Google AI Ultra at $100; Microsoft Autopilot is unpriced. Anthropic bills tokens plus $0.08 per running hour. A do-it-yourself agent costs $0 to $5 a month to host, plus the model.
Part of the Always-on agents series, post 3 of 33.
I pay for agents in three places: the vendor, the machine and the model. The launches this autumn mostly hide the second two inside the first, which is convenient until you want a second agent. So I went through every price page I could find on one day, 20 October 2026, and wrote down what it said. If a figure below is not on a vendor's page, I say where it came from, and if I could not find it at all, I say that too. The companies are described in the headlines post; this one is only about money.
What the hosted agents cost
| Agent | What you pay | What that includes | What is not priced |
|---|---|---|---|
| OpenAI dots | Pro $100, $200 or $500 a month; Business Premium per seat | Your first dot, with a larger allowance in the first month | Every dot after the first; usage terms after month one |
| Meta Muse | Free; Power $20; Maximum $100 a month | Reported weekly allowances of about 100M, 500M and 3B tokens | Per-task limits; small-business pricing |
| Google Gemini Spark | Google AI Ultra, $100 a month (WSJ); a $200 Ultra tier also includes it | Spark in US beta, plus the rest of Ultra | Any plan below Ultra; countries outside the US |
| Microsoft Autopilot | Copilot Credits, usage-based; default off for enterprise tenants | Nothing public yet | The agent itself; credit consumption per task |
| Anthropic Managed Agents | Standard token rates plus $0.08 per running session-hour (as reported) | Runtime, vaults, scheduled runs | Nothing hidden that I found; idle time is not billed |
| Claude Code routines | Included in Pro, Max, Team and Enterprise | 5, 15 or 25 runs a day by plan; more on paid usage | The paid-usage rate per extra run |
| Notion custom agents | Credits, $10 per 1,000 a month, on Business and Enterprise | Agents on a schedule or trigger | How many credits a given task uses |
Two things stand out to me in that table. First, the Business Premium figure for dots is in dispute: CBS News reported $20 a month, a help-centre based write-up reported $125 per user per month or $100 billed annually with a two-seat minimum. I would plan on the higher number. Second, the only vendor that tells you what a running hour costs is Anthropic, and even that figure comes from third-party guides that say they copied the official page, so check it before you budget.
What the free hosting tiers really include
If you run your own agent, the machine can be free, with conditions that changed this year and that most guides still get wrong.
Oracle Cloud Always Free
The Arm allowance was cut on 15 June 2026 from 4 OCPUs and 24 GB of memory to 2 OCPUs and 12 GB, with no blog post or email; Oracle changed the documentation and, from 18 August, began terminating instances over the new limit. You also get two small AMD VMs, 200 GB of block storage and no expiry. The catches are real: Arm capacity in popular regions runs out, so provisioning can fail until you retry, the home region cannot be changed after signup, and idle instances can be reclaimed. Even halved, it is the most capable free box on this list.
Google Cloud
One e2-micro a month, free only in us-west1, us-central1 and us-east1, with 30 GB of standard persistent disk and 1 GB of North American egress. Three details that cost money if you miss them: the console defaults to balanced disk, which is not covered, so pick Standard; an external IPv4 address is $0.005 an hour with one hour free, so a VM that runs all month pays about $3.65 for its address, while IPv6 is free; and the tier requires a billing account. The machine is a shared-core burstable instance with 1 GB of memory, which is enough for an agent that calls a hosted model and not enough for a local one.
AWS, Azure and the rest
The twelve-month EC2 free tier is gone for accounts opened after 15 July 2025. New accounts get up to $200 in credits, $100 at signup and $100 for onboarding steps, and the Free plan closes after six months or when the credits run out. Azure gives a $200 credit for 30 days plus twelve months of small VMs; neither renews. Render offers 750 free instance-hours a month but spins idle services down after 15 minutes, which is the opposite of always-on. Fly.io removed its permanent free allowance in 2024 and Heroku ended free dynos in 2022, so if a guide tells you to use either for free, it is old.
The five-dollar server
For most people I would skip the free tiers and pay for a small VPS. The reason is not capacity, it is that the free tiers come with reclaim rules, capacity errors and region locks that you will hit at 3am, and the Hermes Agent project's own example host is a $5 VPS for a reason. At that price you get a machine that is yours, in a region you choose, with an IPv4 address included. The setup guides later in this series assume this box, and they also work on the free ones.
The model: $0 to whatever you let it spend
A hosted model is the only line in an agent's budget that scales with how much it thinks. The vendors above bundle it; a do-it-yourself agent pays per token. The alternative is a local model through a tool like Ollama, where the API bill is zero and the cost moves into hardware and patience. I cover which small models can actually hold a tool loop together in the open-source section of the series. What I will say here is that one reviewer running OpenClaw with a large hosted model reported about $700 in API costs in roughly three weeks, of which about $600 was one model, and the OpenClaw guides I trust put light use at $5 to $20 a month and heavy use at $100 or more. The spread is the whole point: an always-on agent with an uncapped model budget is an open tab.
What my own agents cost
I will be exact about this, because vague cost claims are how people get burned.
- The publishing agent that puts this post live each morning uses no language model at all. It is a scheduled job on a Mac I already own. Its cloud cost is a handful of storage operations and one cache purge a day, which rounds to zero on my bill.
- The site watchdog that checks the blog at 09:40 is the same: rules, no model, a few HTTP requests and one email when something breaks.
- The schedule shift I ran today, moving fifteen posts and sixteen newsletter sends back five weeks, was a script that asserted every send's status before touching it. No model, no cost, and I would not have let a model near it.
- Leo, the product, is the one that spends. I am not going to publish our internal model spend, and I would be suspicious of any vendor who published theirs without context. What I can tell you is the shape: roughly one model call per candidate clip, once an hour, and the biggest saving we ever made was the rule that no match means no post, because an agent that has to produce something every hour spends on filler.
How I budget an agent before I switch it on
- Price the machine first, then the model. The machine is $0 to $5 and fixed. The model is the variable, and it is the one to cap.
- Put a hard cap on the model, not a soft target. Every hosted model lets you set a monthly limit. I set it before the first run, at a number I would not notice losing, and raise it only after a month of logs.
- Measure cost per run, not per month. A monthly bill hides a loop that quietly doubled. My jobs log every run; a run that costs more than the last ten is the first alarm I look at.
- Count the second agent now. Every consumer launch includes the first agent and prices nothing after it. If your plan has three agents in it, you do not yet know what it costs.
The rest of the series, including the free setup on every system, is on one page. The definition I use for all of this is in the first post.
Sources
- Szymon Paluch, OpenAI dots: ChatGPT's always-on agents, and who can use them (accessed October 20, 2026)
- CBS News, Sam Altman unveils dots, OpenAI's new AI personal agent (accessed October 20, 2026)
- DataCamp, OpenAI dots: always-on agents in ChatGPT, explained (accessed October 20, 2026)
- TechCrunch, Meta debuts its Muse AI agent. Will consumers trust it? (accessed October 20, 2026)
- Tech Insider, Meta launches Muse AI agent: $20 and $100 tiers (accessed October 20, 2026)
- Yahoo Tech, Google launches Gemini Spark personal AI agent at I/O 2026 (accessed October 20, 2026)
- DataCamp, Gemini Spark: Google's always-on AI agent explained (accessed October 20, 2026)
- tbreak, Microsoft Autopilot enters private preview for businesses (accessed October 20, 2026)
- TrueFoundry, Claude Managed Agents pricing (accessed October 20, 2026)
- Tessl, Anthropic adds routines to Claude Code for scheduled agent tasks (accessed October 20, 2026)
- Notion, Custom Agents: your 24/7 AI team (accessed October 20, 2026)
- MrPlanB, 5 best free VPS tiers in 2026 (Oracle just halved its Arm tier) (accessed October 20, 2026)
- Klymentiev, Free VPS in 2026: 6 real options (Oracle, GCP, AWS) (accessed October 20, 2026)
- AgentDeals, GCP free tier complete guide 2026 (accessed October 20, 2026)
- GitHub, NousResearch/hermes-agent (accessed October 20, 2026)
- AI Tool Briefing, OpenClaw review 2026 (accessed October 20, 2026)
- Astropad, How to set up OpenClaw on Mac mini (2026) (accessed October 20, 2026)
Written by Lera Leonteva. Prices are what each linked page showed on the access date and will change; the free-tier rules changed twice this year without announcement.