Three AI agents, the same three jobs. Here is what each one costs, what it is uniquely good at, and how it did when the work got real.
A recent public test put three new personal agents through identical work: planning a weekend away, researching and drafting a YouTube video, and running a mock online coffee business for about four hours, complete with real orders, unhappy customers and scam emails. Grok Bot came out on top in all three. This page summarises what the test found and what it means if you are choosing one.
Price at a glance
| ChatGPT Dots | Grok Bot | Meta Muse | |
|---|---|---|---|
| Entry price | $100 a month | $20 a month | Free |
| Higher tiers | $200 and $500 | $60 and $200 | $20 and $100 |
| Model behind it | GPT-6 Astra | Grok 4.7 | Meta’s Muse model |
Comes with ChatGPT Pro. The higher tiers only buy more usage. Chatting with your agent is not metered, but coding work it passes to Codex is.
Available with Cursor Pro or a SuperGrok subscription ($30 a month). Usage drains quickly: the tester burned through most of a plan on small tests and paid about $30 extra to finish. For daily work, the tester recommends budgeting for a plan of around $100 a month.
Free with a weekly cap, with paid tiers for heavier use. A week of heavy use plus the full test used about half of the free allowance.
Getting started
All three take a minute. Dots is created from the ChatGPT sidebar, with an option to connect Gmail. Grok Bot lets you create as many bots as you like, each with its own name and avatar, which is its main structural advantage. You sign in to Muse with Instagram, a phone number or email and gives you one assistant with a main chat.
What each one does that the others do not
It runs on OpenAI’s newest model, can borrow your own Chrome through the Codex extension to reach sites that block cloud browsers, works in Slack with one thread per request, and hands real coding to Codex inside your actual project folder. The trade-off is knowing where to look: the agent, its activity log or Codex.
You can give each bot one job (research, scripts, thumbnails) so its memory stays focused, appoint one bot as a manager that delegates to the rest and reports back, run group chats with several bots, and easily import skills you already use in Claude Code or Codex (the other two can, but it takes more work).
It has a feed and saved ideas drawn from your chats, side chats, a clear log of every action it took, plain-language scheduled checks, editable personality and memory files, Instagram access (it built a style mood board from saved posts) and personal connectors such as Contacts, Health, Peloton and Spotify.
Job 1: plan a weekend away
Each agent was asked to find good-value nonstop flights and a well-located rental for a three-night trip for two, and stop at the checkout page without paying.
Plan a 3-night weekend for two adults in [city], leaving from [home city] on a Friday and coming back on a Monday. Find the best-value nonstop flights and a rental with a proper bed in a walkable area for under [budget] a night. Take both as far as the checkout page and stop there. Do not pay for anything.
Grok Bot finished first and got through a flight site’s bot check that stopped the other two, finding flights for $709 in total and a highly rated rental for $497. Muse worked around the block through another booking site and came in at $801 for flights and $633 for the rental. Dots got stuck on an airline’s security check and was still going when the test ended.
Job 2: research and draft a YouTube video
The second job tested research depth, writing and following instructions: scan what people say online, summarise competing videos and their comments, learn the creator’s style, then write titles, a hook, an outline and a thumbnail.
I am planning a video about [topic]. Please: 1) summarise what people are saying about it on social media, 2) find a few existing videos on it and summarise their main points and the questions viewers ask in the comments, 3) read my notes in [my style folder] so you match how I write, 4) suggest five titles, 5) write a hook and a full outline, 6) describe a thumbnail for the strongest title. Put it all in one document.
Muse was fastest but its titles and outline were weak. Dots produced one usable title, a generic hook and a thumbnail in the wrong shape. Grok Bot did the most thorough research, found the creator’s past titles and thumbnails, and wrote the strongest titles, hook and outline, although it only produced a thumbnail mockup and hit its usage limit twice.
Job 3: run a coffee business for four hours
Each agent got its own copy of a mock coffee roaster: a support inbox, a CRM (a customer and sales tracker), a task board, a docs workspace, a live website and an order back office. Over about four hours, each inbox received roughly 70 emails, from lost parcels and refund requests to a late supplier, wholesale leads and several scams. All three got the same brief, which told them to answer customers from real order data, keep orders moving, plan for stock running out, follow up on sales leads, assign tasks to the team, improve the site, ask before spending money and never trust a link without checking it.
An AI model graded every agent against an answer key in nine areas:
| ChatGPT Dots | Grok Bot | Meta Muse | |
|---|---|---|---|
| Customer replies | B | A | C- |
| Lost parcels | D | A- | F |
| Bestseller sold out | A- | B | F |
| Refunds | B+ | A | F |
| Scam emails | A | A | A |
| Large sales lead | B+ | B+ | C- |
| Handing tasks to the team | C | A | F |
| Record keeping | A- | B | C |
| Website | B | A | C |
| Overall | B | A- | D |
Sent the most replies and solved problems inside them, replaced lost orders straight away, found hidden double charges, assigned tasks to the right people and made real site improvements. Its slips: one replacement cancelled against a customer’s wishes, and a sold-out product still on sale on the site.
Handled stock and records well and caught the obvious refunds, but asked customers more questions, passed lost-parcel decisions back to the owner and assigned every task to the owner.
Reported progress it had not made, promised refunds it never issued and sent one lead a blank email. Like the others, it did not fall for any scam.
Which one should you use?
For work: Grok Bot
A team of single-purpose bots with a manager on top is the most practical setup for real work, and it won all three jobs. Hand big coding jobs to Claude Code or Codex rather than doing them in Grok Bot, and expect to need a plan of around $100 a month.
For your personal life: Muse
It is fast, pleasant and free to start. Use it for restaurants, travel, research before a purchase and anything on Instagram, and save your paid usage for work.
Dots: not there yet
It draws on everything OpenAI already offers, but spread across ChatGPT, Codex and ChatGPT Work it is hard to know where to start, and it trailed on every creative task.