Grok Bot is built around persistent named teammates: Bots that keep memory, files, browser sessions and preferences across turns on one shared cloud Linux VM per account, with learned routines and computer use for apps that have no API. OpenAI's agent surface, which the alternatives lists name ChatGPT Work in its business packaging, is organised around task sessions in per-task sandboxes. Persistence buys context that compounds and costs you a growing blast radius, because xAI documents that Bots share a computer without separate security boundaries. Disposable sandboxes buy a narrow blast radius and cost you the context. Neither is strictly better, and for developers the more useful third answer is a workbench that keeps coding work on hardware you own.
- Grok Bot persists; the ChatGPT agent surface is session-shaped. That is the architectural difference everything else follows from.
- Grok Bot shares one cloud computer across every Bot on the account, with no separate security boundaries, per xAI's own documentation.
- A per-session sandbox is a smaller blast radius and a worse memory. You cannot have both properties at once.
- Grok Bot needs a Cursor account on every access path, which is an unusual requirement for a non-coding product.
- Neither vendor publishes everything. xAI's pricing page returns 403 to us; check OpenAI's own pricing page rather than a third-party table.
- For repository work, neither is the right shape. Coding agents want worktrees, diffs and a plan gate.
First, the naming
Three separate naming problems collide in this query, and untangling them is most of the value of the page.
| What is typed | What is usually meant | What it is not |
|---|---|---|
| grok bot | xAI's always-on agent product, launched 11 August 2026 | Not the Grok chatbot, and not a bot you build on the Grok API |
| chatgpt agent | OpenAI's agentic surface inside ChatGPT | Not the Assistants API and not a custom GPT |
| chatgpt work | The business packaging of that agent surface, as the alternatives lists name it | Not a separate model |
| grok bot api | Usually people wanting to automate Bots | There is no public Grok Bot API. Only Grok Build and the xAI API are programmable |
| grok agent mode | Split across two unrelated products | Imagine Agent Mode is a creative canvas; Grok Build is the coding CLI |
For the rest of this page, "the ChatGPT agent" means OpenAI's agentic surface as a category, and the comparison is architectural rather than a line-by-line feature audit. We are deliberately not restating OpenAI plan prices or seat counts from third-party tables, because those change and the tables ranking today are frequently stale. Check OpenAI's own pricing page for current numbers, in the same way you should check xAI's.
Persistent teammates against task sessions
Grok Bot's central design decision is that a Bot is a continuing entity. xAI's documentation states that named Bots keep memory, files, browser sessions and preferences across turns, and describes context that "compounds instead of resetting to a fresh environment on every task". You create a Bot, you name it, and over weeks it accumulates the logins, files and habits that make it useful for your particular work. Bots can message each other, share context in threads or group chats, and hand ownership of a task between themselves.
The ChatGPT agent surface makes the opposite bet. The unit is a task: you describe what you want, the agent works in a sandboxed environment for that task, and the environment does not persist as a long-lived shared machine full of accumulated state. Continuity comes from the conversation and from whatever memory the product keeps, not from a computer that stays around.
| Property | Persistent teammate (Grok Bot) | Task session (ChatGPT agent) |
|---|---|---|
| Context across days | Accumulates | Mostly starts fresh per task |
| Logins and sessions | Persist, and are shared across all your Bots | Scoped to the task session |
| Blast radius if one run goes wrong | Everything on the shared computer | What that session could reach |
| Teaching a procedure | Demonstrate once, saved as a routine | Re-describe, or build it as a saved workflow |
| Scheduled recurring work | Routines re-run on a schedule | Depends on the surface you wire it to |
| Parallel agents | Several Bots, own screens, one filesystem | Separate task sessions |
| Cleanup | Deleting a Bot is reported not to remove shared files or sessions | The session ends |
Read the third and last rows together, because they are the same property seen from two angles. Persistence is what makes a Bot feel like a colleague rather than a vending machine, and it is also what makes the thing you cannot easily clean up. xAI is upfront about the first half. The second half is documented obliquely: deleting a Bot does not remove the shared files or browser sessions it used, which are reported to require several manual sign-out steps.
The trust difference, stated plainly
This is the part of the comparison that should decide it for anybody connecting an agent to something with money attached, and it is the part the vendor comparisons skip.
xAI's documentation says three things that need reading together. All of your Bots use the same persistent cloud computer. Each Bot gets its own screen on that computer, "so several Bots can use browser and desktop tools in parallel without getting separate security boundaries". And: "Treat a login or file placed on the computer as available to all of your Bots."
That is not a security failure. It is a stated design choice with a stated consequence, and xAI deserves credit for stating it, because plenty of vendors would not have. But it means a Bot you created for a low-stakes task can reach the session you established for a high-stakes one. A per-task sandbox model does not have that property, and that is its main advantage.
- Approvals are prose, not policy. Grok Bot's guardrails are written instructions such as "never send external messages without approval", with an Auto Review layer that is a model judging a model. That is a useful safety net and it is not an enforced permission system.
- There is no dry-run mode. eesel's review states it directly: a test run performs real work, navigating sites, changing files and calling tools. Rehearsing a send means you sent it.
- There is no queryable organisation-wide audit view. The documentation describes an audit view as coming, twice. Today there are per-Bot chat transcripts, and the twenty most recent routine records are reported by Daily Dose of Data Science and unconfirmed by xAI.
- No published compliance claim was found in the doc set. No published SOC 2, ISO 27001, GDPR, or HIPAA claim was found, nor a retention period or data-residency option. Compliance defers to Cursor's terms rather than xAI's own.
- Cursor Legacy Privacy Mode blocks Grok Bot entirely. Data storage is mandatory, so a team that adopted Cursor on the condition of privacy mode cannot use Grok Bot at all.
The accountability point raised repeatedly in the Hacker News launch thread belongs here too. Commenter SilverBirch framed it as a question about identity: a Bot that holds your credentials and acts as you on the web produces logs that say it was you. That is true of any agent inheriting your sessions, on any vendor, and it is a reason to scope what an agent can reach rather than a reason to avoid the category. It is also a reason to prefer session-scoped access over a computer that quietly collects every login you have ever handed it.
What each one costs you, in shape rather than digits
Both products bundle rather than meter transparently, and both leave you unable to answer "what did that task cost" with precision. The shapes differ enough to matter.
| Grok Bot | ChatGPT agent | |
|---|---|---|
| How you buy it | Bundled into a Cursor or SuperGrok subscription, Cursor account required in every case | Bundled into a ChatGPT plan or business packaging |
| Cheapest eligible path | Cursor Teams Standard at $40 per seat, or Cursor Pro+ at $60 | Check OpenAI's current pricing page |
| Known published prices | Cursor Ultra $200, Cursor Teams Premium $120 per seat | Published by OpenAI |
| Prices we could not confirm | SuperGrok Plus and SuperGrok Heavy, reported at $100 and $300 | None known |
| Metering | Undisclosed weekly token allowance, no product-level spend cap reported | Plan-scoped |
| Model choice | None. Automatic routing, no picker | Model selection available on the ChatGPT surface |
| Free trial | Announced with limited usage; xAI states no duration | Check OpenAI directly |
Two rows deserve calling out. First, the Cursor account requirement is genuinely unusual for a product aimed at office work: a marketing operations lead evaluating Grok Bot for inbox triage is going to have to create an account with a developer-tools company to do it. Second, the free trial. xAI announced a free trial with limited usage for users outside the eligible plans alongside the 21 August access expansion, and has published no duration for it anywhere we could find. Third-party pages claiming 7 days or 14 days are not citing an xAI source, and at least two of them disagree with each other.
What to verify before you connect either one
Both vendors will happily take a login on the first run. Six things to establish first, whichever you pick, because none of them is reversible after an agent has acted.
Use a scoped account, not your own
Create a service account with access to exactly what the task needs. This is the single highest-value control in the whole category, and it is the one that survives being wrong about everything else. On Grok Bot it matters more than usual, because a session established for one Bot is available to all of them.
Establish what a mistake costs
Neither product offers a genuine dry run. On Grok Bot, eesel documents that a test run performs real work: it navigates real websites, changes real files and calls real tools. Rehearse against a sandbox tenant, not production, and assume the first run is the real one.
Find out what you can prove afterwards
Grok Bot has per-Bot chat transcripts rather than a queryable organisation-wide audit view, with an audit view documented as coming. The twenty most recent routine records are reported by Daily Dose of Data Science and unconfirmed by xAI. If you will need to reconstruct what happened for a customer or an auditor, check this before you scale usage rather than after.
Read the prompt-injection position honestly
The Hacker News launch thread spent most of its length here, and the strongest point in it was statistical rather than rhetorical: commenter stymaar argued that a system failing every fortieth attempt is catastrophic rather than tolerable. Assume any agent that reads untrusted web content can be steered by it, and scope the account accordingly.
Check the compliance surface, not the marketing page
No published SOC 2, ISO 27001, GDPR, or HIPAA claim was found in the Grok Bot documentation set, there is no stated retention period and no data-residency option, and compliance defers to Cursor's terms. That may be fine for your use and it is not fine for everybody's.
Know how to revoke access
On Grok Bot, deleting a Bot is reported not to remove the shared files or browser sessions it used, which require several manual sign-out steps. Work out the revocation path while nothing is wrong.
None of this is exotic. It is the same list you would work through before giving a contractor a laptop, and the reason it feels heavier here is that the contractor is faster and never gets tired.
Which one should you actually buy
Four scenarios, and they resolve cleanly.
Repeated procedures across apps that have no API
Grok Bot. Learn-by-demonstration routines plus computer use plus persistent browser sessions is a combination nothing else assembles as neatly, and it is the reason the product exists.
One-off research, analysis and document work
The ChatGPT agent surface. A task session is the right shape for a task, and you get the narrower blast radius for free.
Anything touching money, customer data or production systems
The session-scoped model, and a scoped account rather than your own login, on either vendor. Grok Bot's shared computer plus missing audit log plus absent dry-run is a hard combination to take to a security review today.
Repository work
Neither. Coding agents want branch-level isolation, a diff you can review hunk by hunk, and a plan gate before writes. A general teammate on a shared cloud filesystem provides none of those.
That last one is worth expanding, because a lot of the people searching this comparison are developers. If the work is code, the useful question is not which cloud teammate to rent but where the agents should run. Continuum runs Claude Code, Codex, Cursor, Grok, Gemini, and OpenCode on a Mac, Linux or Windows host you already own, gives each Code session its own git worktree, gates writes behind a plan you approve, records spend by repository and provider, and is driven from a native iPhone client or the web so the host keeps working while you are somewhere else. The app is free. The repository and credentials remain on your host; selected context is sent to the configured provider for inference.
For the wider field, the alternatives guide ranks eight options including OpenClaw, Hermes Agent and Manus. For the naming maze around xAI's own products, Grok Bot versus Grok Build resolves which of them is actually programmable.
Questions people ask
What is the difference between Grok Bot and ChatGPT agent?
Grok Bot is built around persistent named teammates that keep memory, files and browser sessions across turns on one shared cloud computer per account. OpenAI's agent surface is organised around task sessions in per-task sandboxes. Persistence buys accumulating context and costs you a growing blast radius; disposable sandboxes buy a narrow blast radius and cost you the context.
Is Grok Bot the same as ChatGPT Work?
No. ChatGPT Work is how the alternatives lists name OpenAI's agent surface in its business packaging. Grok Bot is SpaceXAI's product, launched 11 August 2026, and it requires a Cursor account plus an eligible SuperGrok or Cursor subscription on every access path.
Which is more secure, Grok Bot or a ChatGPT agent session?
On isolation, the session-scoped model, for the reason set out above: a task sandbox is discarded while the shared Bot computer accumulates logins across every Bot you run. Grok Bot also has no dry-run mode, no queryable organisation-wide audit view, and no published compliance claim found in its documentation set.
Does Grok Bot need a Cursor account?
Yes, on every access path. The documentation states that Grok Bot uses your Cursor account and that you sign in with it, whether your eligible plan is SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra, or Cursor Teams Standard or Premium.
Can Grok Bot or a ChatGPT agent replace a coding agent?
For small self-contained tasks, sometimes. For repository work, no. Neither provides branch-level worktree isolation, a hunk-by-hunk diff review, a plan gate before writes, or per-repository cost attribution. Those are the features a coding harness exists to provide.
Which model does Grok Bot run?
xAI does not disclose it. There is no model picker and routing is automatic. Grok 4.6 shipped on 12 August 2026 and is available in Grok Bot among other surfaces, so it is the plausible default, but no xAI page names a default and you cannot control which model handled a task or what it cost.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.
- Grok Bot overview (docs.x.ai) shared cloud computer, security-boundary warning, memory, routines, connectors and computer use
- Grok Bot get started (docs.x.ai) eligible plans, Cursor account requirement, platform support, Linux statement
- Grok Bot on more plans (xAI) 21 August 2026 access expansion to SuperGrok Plus, Cursor Pro+, and all Cursor Teams plans
- Grok Bot review (eesel) compliance gaps, missing audit log, no dry-run mode, metering behaviour
- Grok Bot on Hacker News launch-day thread, 350 points, trust, prompt injection and self-hosting objections
- Grok 4.6 (docs.x.ai) context window, knowledge cutoff, reasoning effort, surfaces it ships on
- Continuum pricing Free, Plus, Max 100, Max 200 and Ultra plans and their weekly hosted-usage allowances