Grok Bot vs OpenAI Dots: Which Fits Your Recurring Work?

Documentation comparison, checked October 9, 2026. We have not run a controlled comparison inside the two products. Recommendations below concern documented workflow fit, not measured speed, quality or task cost.

Grok Bot and OpenAI Dots are worth comparing when your goal is to hand off a project that continues beyond one conversation. The useful question is how each product organizes work, retains context, reaches your tools and handles recurring tasks. A model leaderboard does not answer those questions.

The clearest documented distinction is work organization. Grok Bot supports several named Bots with separate conversational context on a shared account computer. Dots starts with a primary persistent assistant that can draw on ChatGPT memory. Both describe ongoing work and scheduling. Neither description proves that a particular account can access every advertised feature, or that a completed task will cost less. This article separates documented features, our editorial selection criteria and practical acceptance criteria for your own workflow. Grok Bot overview, Introducing dots.

Compare the agent products before comparing their models

An API request and a background agent have different responsibilities. An API integration sends a defined request and processes a response. An agent project may also need to find a source, authenticate to an application, operate a browser, save files, retain context and return later. A failure at any of those stages can make the final deliverable unusable even when its prose looks convincing.

For this comparison, a useful unit of work is an accepted deliverable: a sourced research brief, an editable content package or a scheduled check with an actual execution record. A polished chat response is insufficient if the requested files never arrive. Conversely, a login interruption should be recorded as an access limitation rather than automatically attributed to the language model.

This distinction matters to Ofox readers who already use APIs for production workflows. A Grok model endpoint does not by itself give an application the Grok Bot product. Likewise, access to an OpenAI model is not evidence of Dots account eligibility. Do not select an agent subscription by substituting a model’s token price for the cost of completing the whole job.

The documented differences that affect your setup

These are documentation differences, not measured advantages. The table intentionally does not score speed, answer quality or reliability. Sources: Grok overview, skills and routines, Grok plans, Dots introduction, Dots help.

Start with account eligibility, not another subscription

Before paying for anything, identify the exact account you intend to use and check its existing entitlement. Grok’s plans documentation lists access through qualifying Cursor plans and linked Grok or X subscriptions, with business and enterprise routes handled separately. An additional eligible subscription does not necessarily provide additive usage. Check the account pairing carefully rather than treating a second purchase as a quota refill. Grok plans and billing.

Dots eligibility: Pro rollout excludes the EEA, Switzerland and UK; Business Premium covers supported ChatGPT regions; Enterprise beta requires administrator enablement. Rollout can still delay individual access. Check desktop rather than treating a launch announcement as account entitlement. Getting started with your dot.

For a team evaluation, write down the plan, account region, client version and any administrator restrictions before the first task. Keep this private record separate from public screenshots. If one product has a connected source and the other does not, record that difference before interpreting the output. Otherwise you risk attributing an information-access advantage to reasoning ability.

Separate roles do not automatically isolate files

Multiple named assistants can make ownership easier to understand. For example, an editorial team could give one role responsibility for source collection and another responsibility for adapting an approved interview. That is an organizational choice, not a result from our tests.

Grok’s documentation explicitly describes a shared computer across the account’s Bots. Separate conversations should therefore not be treated as separate file-security boundaries. Grok Bot overview.

Use task-specific folders and explicit file permissions in your operating process. A useful handoff names the input, the permitted destination and the acceptance criteria. “Improve the launch content” leaves much more room for misunderstanding than “Read this approved transcript; write three internal drafts in this output folder; preserve the source unchanged.”

For Dots, existing ChatGPT memory may provide useful background, but it can also make a comparison asymmetric if the other product starts with less context. Record this rather than deleting the user’s memories to manufacture a clean test. Ask either assistant to verify current facts against the supplied source. Familiarity with a project should not override an updated decision in the latest interview.

Connecting an app is different from signing into its website

Grok documents two access routes: a structured connector for a supported service and browser interaction for other workflows. A browser login and a connector authorization are different states. Its installed connectors are account-wide, so naming a separate Bot is not enough to restrict their availability. Use the computer and apps.

Dots uses ChatGPT app controls. Connected information may also inform proactive work; disconnecting an app does not erase information already obtained. Local-computer access is optional and initially off. These distinctions affect what you expose during a trial. Dots help.

For either product, validate the smallest required operation before granting a larger job. If the task is to summarize a document, ask it to read that specific document and identify its version. Do not authorize an entire publishing workflow merely to prove that a connector is installed. Record the authorized account, accessible source and permitted action separately. This makes a later failure diagnosable: the wrong account, expired authorization and an unsupported operation are different problems.

A repeatable task needs more than a reusable prompt

For a recurring brief, define four things before scheduling: the source boundary, what counts as new, where the result belongs and what happens when a source fails. A daily report that silently reuses yesterday’s material may look successful while delivering no fresh information.

Grok documents a distinction between a reusable skill and the routine that determines when it runs. Dots also supports recurring checks. Neither should be reduced to a label such as “has automation” without verifying the whole workflow. Grok skills and routines, Dots scheduled tasks.

For a first evaluation, use one scheduled execution with a harmless local input. Capture the task identifier, enabled state, time zone and scheduled time. Then capture the actual run and its result. A message saying “I’ll do that tomorrow” proves neither registration nor execution. A manual test run can validate parts of the task, but it cannot demonstrate that the scheduled trigger worked.

There is also a version-specific documentation trap. Grok’s October 7 changelog says typing / no longer opens the skills-and-actions menu, while its skills page still describes that menu. The October 2 release renamed Marketplace to Connect Apps. Use the installed version when following screenshots; a missing old menu is not enough to diagnose broken skills. Grok Bot changelog.

Measure cost per accepted output with the right units

Subscription price, included usage and extra charges answer different questions. For Grok, the official billing page distinguishes weekly included usage from monthly on-demand consumption. It also warns that a running task can finish beyond the monthly limit. Do not describe that limit as an absolute mid-task spending stop. Plans and billing.

Record usage before and after each task in the units the product actually displays. Missing information stays unknown; it does not become zero. Currency, usage percentage, credits and task counts cannot be mixed into one apparent cost number without a documented conversion.

For Dots, the first dot is included with Pro or Business Premium, alongside a deeper-work allowance. The launch page distinguishes conversations with a dot from tasks it starts or manages in Codex or ChatGPT Work: those tasks still use their usual limits. Continuous availability therefore should not be described as unlimited execution. Introducing dots.

For your own purchasing decision, calculate two separate quantities. Incremental cash cost is the extra amount charged for the task. Allocated subscription cost is the share of an existing subscription you choose to attribute to it. Both can be useful, but they are not interchangeable. A task completed within an already-paid allowance may have no visible incremental charge while still using scarce capacity.

Also record the time spent checking and repairing the result. A low bill is not a practical advantage if the content needs extensive correction before use. With a small test set, report task-specific observations rather than declaring one product generally cheaper.

A reusable acceptance brief for either product

The following is an original template, not a transcript of a completed Grok Bot or Dots run. It is useful when you want to judge an assistant by an inspectable result instead of a confident response. Start with material you are allowed to share; a synthetic interview is enough for initial setup.

Read only the attached interview. Create summary.md, posts.md and actions.json.
The summary must distinguish shipped work, future plans and unresolved decisions.
Write three internal post drafts; do not publish or send them.
For each action, record action, owner, deadline and supporting source lines.
Use null for an owner or deadline the interview does not establish.
Cite source lines for factual claims. Preserve the input file unchanged.
Do not invent prices, launch dates, quotes, performance improvements or signup links.
Return the three files and a list of unresolved facts.
If the input cannot be read, report that failure and stop.

This contract exposes several practical differences that a feature table cannot predict. Does the assistant deliver actual editable files? Do the citations identify the supplied material? Does a discussed but rejected price accidentally become the advertised price? Does a planned feature become a shipped feature? A correct JSON shape cannot answer those semantic questions.

Review the original output before revising it. Keep the source, original files and edited versions separate. Count concrete corrections, such as an invented deadline or omitted limitation, rather than assigning an unexplained quality score. If only one product has been tried, describe that experience without implying a comparison result.

For recurring research, add a second-run rule: compare current sources with the previous accepted claims file, report substantive changes only, and identify sources that failed. A changed retrieval timestamp is not a product update. Save the fetched source versions so a later reviewer can distinguish changed evidence from changed interpretation.

For a scheduled check, request one harmless run before establishing a recurring workflow. Record registration and execution separately, including time zone. Preserve the run record if the result is missing. A manual test can check the instructions, but only the actual scheduled execution tests the trigger. Choose an observation window in advance and report what happened within that window without generalizing a single run into reliability statistics.

Which should you evaluate first?

Based on the documented product designs, Grok Bot is a reasonable first candidate if your main requirement is several persistent named roles sharing a working environment. Dots is a reasonable first candidate if continuity with your existing ChatGPT context is central to the work. These are editorial fit judgments, not measured performance findings.

For research, content adaptation and recurring checks, prioritize whichever product you can already access with the necessary sources and a clear consumption boundary. Run the smallest complete job before expanding the scope. A second product is worth testing when the first leaves a specific gap, such as poor source traceability or costly manual correction—not simply because a launch announcement lists more features.

Migration should preserve the task contract rather than copy an old conversation blindly. Transfer approved source files, output examples, decision rules and acceptance checks. Verify permissions and schedules afresh. Never assume a connector, remembered preference or scheduled task moves with the text of a prompt.

Related API and computer-use workflows

For a separate API implementation, see our offline computer-use controller. For a concrete research deliverable, see the competitor research table workflow. These are separate workflows, not Grok Bot or Dots integrations.