Is Grok Bot Down? Four Failures That Look Like an Outage, and What to Check First

If your Grok Bot stopped answering this morning, the odds are against it being a platform outage, and the reason is structural. Grok Bot is SpaceXAI's always-on agent, launched on 11 August 2026, and it runs as a small fleet of named bots on a cloud computer assigned to your account. Almost every way it fails is a way that produces the same symptom — silence — while the platform itself is perfectly healthy. So the honest first answer to "is it down?" is that you cannot tell from the symptom alone, and there is an order in which to check that gets you to the answer in about three minutes.

One caveat before the checklist: SpaceXAI's Bot documentation, as we read it, gives you an account-level usage readout and per-routine run records, but no public uptime or incident feed. There is no page you can refresh to learn whether the service is degraded. That absence is the reason this article exists rather than a link to a status page.

The four failures, in the order they actually happen

One: the account ran out of included usage. The most common, and the one that mimics an outage most convincingly, because a bot that has hit its ceiling simply stops. Grok Bot's access is bundled into paid Cursor plans and SuperGrok tiers, and the meter the documentation describes is weekly included usage on the account — not per bot. That distinction is the whole trap: a single chatty bot can exhaust the allowance that five bots share, and from the outside it looks like the fleet died at once. Usage & Billing shows weekly included usage and on-demand usage for eligible accounts, with a weekly readout in the account menu, and a plan that exhausts its included usage can add on-demand spending billed from model and token cost. One line in the same FAQ is worth knowing if you pay for both: if you hold a Cursor subscription and a SuperGrok subscription, the Bot draws on whichever has more usage.

Two: the routine is paused, not broken. Routines are how a bot runs on a schedule or after an event, and they are not guaranteed to keep running through your absence. The documentation notes that Grok Bot may ask whether to keep routines running after a long absence and pause them if you do not answer. A bot that went quiet after a holiday, a long weekend, or an unread prompt is most likely sitting in exactly that state.

Three: the computer was reset and took your installs with it. Every bot on your account shares one computer, with a persistent Linux surface, a browser, a command line and a shared workspace at code>/workspace/code>. The documentation is explicit that temporary directories, manually installed packages and uncommitted application state should be treated as replaceable — and that the two recovery paths differ: Reset restores the last saved snapshot, while Recreate "removes apps and packages members installed themselves." If your bot depends on something you installed by hand — a third-party coding agent, a CLI, a library — a recovery can remove it and leave a bot that answers but cannot do its job. That reads as partial outage and is actually a missing binary.

Four: the model behind the bot changed, or degraded. Cursor manages model selection. The documentation's answer to "where do I pick between Grok 4.7 and another model" is flat: there is no model picker, no per-bot setting, no per-routine one, and no per-skill one. On 7 October, Elon Musk said publicly that Grok Bot would stop being limited to SpaceXAI's own models and would reach for "the best back end model for any given task" — a statement with no rate card, no changelog entry and no outside measurement attached. So the engine is a platform-side decision you cannot see. When it changes, what you observe is not an error message; it is output quality, latency and cost drifting on the same tasks. A bot that suddenly writes worse or takes longer has not necessarily gone down.

The three-minute check

Run these in order and stop at the first one that explains what you see:

• Account usage first. Open the account menu and read the weekly included usage figure before you touch anything else. If it is exhausted or near exhaustion, everything downstream — every bot, every routine — is failing for one reason and no amount of debugging the bot will help. On-demand spending is the lever, and it is a billing decision, not a technical one.

• Run records second. Each routine keeps the 20 most recent run records. That is a deliberately shallow audit trail, and it is still enough to answer the question that matters: did the routine run and fail, or did it never start? A routine with no recent records is paused or never triggered; a routine with records that end in failure is a bot problem; a routine with successful records and no visible output is an output-delivery problem.

• The computer third. Ask the bot directly what it can see — code>/workspace/code> contents, whether an installed tool still exists, whether a sign-in has survived. Because the computer is one machine for the whole account and files are visible to every bot, this also tells you whether something a different bot did has affected this one.

• The trigger fourth. If the bot runs on events rather than a schedule, check the trigger's breadth. The documentation warns against broad listeners such as "every new message," which generate noise, burn usage and act on irrelevant input. A bot that fires constantly and usefully never, or the reverse, is usually a trigger-scoping problem rather than a service problem.

Only after all four are clean is "the platform" a reasonable hypothesis — and even then, the evidence you have is circumstantial, because there is no status feed to confirm it against.

Why the shared computer makes this harder than it should be

The single-computer design is the most consequential thing about Grok Bot and the least obvious until something goes wrong. One computer is assigned to your account rather than to any individual bot; the screens each bot sees are described as "separate work surfaces, not separate security boundaries"; any secret can be used in any teammate's conversation on a Team Bot. That has two consequences for diagnosis.

First, a failure is rarely local. A file written by your research bot is readable by your outreach bot, and a package removed during a reset affects every bot that used it. Second, recovery is account-scoped. Deleting a bot removes the active profile, conversation and routines, while computer files and sign-ins are not isolated per bot and may persist. Hiding a bot is the reversible alternative; deleting a routine is immediate and has no undo.

If you are debugging at 2 AM, that asymmetry matters more than any log: the operations that look like cleanup are the ones with no way back.

What "status" would actually require

A real status answer needs three things Grok Bot does not currently give you, and it is worth naming them because they are the difference between a product you can operate and one you can only watch:

• A per-bot meter. Usage is measured at the account level, so a fleet that fails together cannot be decomposed into which bot caused it.

• A deeper run history. Twenty records per routine is enough to catch a failure this week and not enough to catch drift across a month — the failure mode that costs money quietly rather than loudly.

• An incident feed. Without one, "the service is degraded" is an inference from four absences rather than an observation.

None of that is a criticism of the product's design, which is coherent for what it is: a subscription agent you configure by talking to it. It does mean the burden of proof sits with you every time the bot goes quiet, and the check order above is the cheapest way to carry it.

Where a router fits, and where it does not

Grok Bot is a subscription product, not something you can route through a third-party endpoint, and no honest guide should suggest otherwise. The question the model-selection problem raises is different: if you need to choose the engine on purpose — for cost, for a licence, for reproducibility of a result — that requirement cannot be expressed inside the bot at all. It has to live in the layer below.

That layer is what we sell. OrcaRouter is one OpenAI-compatible endpoint across 200-plus models with no markup on provider list price, so a vendor's price cut is live the same day. Grok 4.6 sits on it at SpaceXAI's own published rate of $2.00 per million input tokens and $6.00 per million output tokens, with a 500,000-token context window and a measured p50 time to first token of 1.65 seconds. The pieces that matter when an unproven step fails are the ones that stop a failure from becoming an incident: automatic failover across providers when one degrades, a routing DSL that composes several models into one call, and model fusion for panels that answer together.

The practical move for anyone reading this because their bot went quiet is the boring one. Write the skill, run it once by hand against the model you were going to trust, and only then automate it. A bot that has been validated against a model you can call directly is a bot whose failures you can attribute — and attribution, not uptime, is what "is it down?" is really asking.