Grok Bot vs Self-Hosted OpenClaw: The $120 Question

Bottom Line

Grok Bot is the first agent product from a major lab that matches what OpenClaw users already do: agents that persist, hold memory, use a real computer, and keep working when you close the laptop.

It is also, by the account of its own early users, hard to predict the cost of.

The trade is clean. Hosted removes the hardware problem and adds a metering problem. Self-hosted keeps the hardware problem and removes the metering problem. Which one you want depends on whether your agents run in bursts or continuously.

What Grok Bot actually is

xAI opened the beta on August 11, 2026. Each bot gets:

• a dedicated computer environment in the cloud
• persistent memory across sessions
• tools, and the ability to operate the apps you already use
• coordination with other bots
• execution that continues while you are offline

Access is bundled into a paid plan at roughly $120 per month rather than sold on its own.

Early access users describe it as less like prompting an agent and more like handing work to a capable teammate. That framing is the product’s real advance — not any single capability, but the fact that non-technical operators immediately understand what it is for.

The complaint everyone has

Quota burn, and it is not a minor grumble.

• One user reported consuming 42% of a weekly allowance on day one.
• Another measured roughly 100 chat completions plus a single 10-minute script consuming 5% of the week, and called it unusable for real business work.
• Even trivial requests were reportedly routed through the heavyweight model tier.

The structural reasons reported alongside it matter more than the numbers:

• no published weekly allowance — you cannot plan against a limit you cannot see
• no spend cap
• no model picker — you cannot send cheap work to a cheap model
• billing follows whichever model happened to serve you

Users also report daily Chrome profile resets, stalled agents, and no way to prune a bot’s ever-growing context. Security commentary has focused on a gap between the friendly team metaphor and an account-level machine underneath: multiple bots do not isolate their access from one another.

What that costs you against a local rig

This is the comparison this site exists to make.

On a self-hosted OpenClaw setup running local weights, the marginal cost of a long agent run is electricity. An agent that thinks for an hour costs the same as an agent that thinks for a minute, plus a few cents of power. That is the entire argument for local inference and it does not change because someone shipped a nicer product.

On Grok Bot, that same hour consumes an allowance you cannot see, at a rate set by a model you did not pick.

Neither is strictly better. But they fail in opposite directions:

If your agent work is bursty and light, the hosted option is obviously simpler and the quota may never bite. If you run agents continuously — the exact case that makes local hardware pay for itself — an unpublished weekly allowance is the wrong shape of cost.

Who the reviews say should wait

Early reviews are unusually consistent about the fit:

Reasonable fit today: non-technical operators, draft-first work, anything reversible.

Poor fit today: heavy multi-agent operations, schedule-critical automation, regulated work, and experienced agent builders.

The recurring summary is worth quoting in spirit: worth trying, not yet worth reorganising a business around.

That last group is most of this site’s readers. If you already run OpenClaw, you are in the category the reviews explicitly say to wait on — not because Grok Bot is bad, but because you have already solved the problem it solves, in a way that bills you in watts rather than tokens.

What would change the calculation

Three things, and they are all things xAI could ship:

1. A published allowance and a spend cap. Most of the complaint disappears the moment you can see the meter.
2. A model picker. Routing cheap work to a cheap tier is the single biggest cost lever in any agent stack.
3. Real isolation between bots. The team metaphor only holds if one bot’s access is not every bot’s access.

None of these are hard. It is a beta three weeks old. We will revisit this when they land.

What we have not tested

We have not run Grok Bot ourselves. Everything above is drawn from xAI’s own announcement and from published early-user reports, and it reflects a beta that is weeks old and changing.

What we can speak to is the other side of the comparison: the cost curve of running agents on local hardware, which is the subject of most of this site. If you want the arithmetic for your own workload, the gear ladder lists what each tier actually costs to buy and run.