Marko

Building AI agents that do real work.
Setups, costs, receipts

[Marko @Marko_Poly]

https://t.co/G0F7SXoClD

[Marko @Marko_Poly]

@davidim Same controls at the Seedance API price is the kind of math people actually switch for. Congrats on the launch.

[Marko @Marko_Poly]

Same cinematic controls. About half the price per clip. #ad

Sume just launched Director Studio 1.0. It works like Higgsfield Cinema Studio: camera, lens, aperture, genre, era, tempo, lighting, emotion, an AI cast and up to 50 references.

The difference is the bill. Sume charges the Seedance 2.5 API price with 0% markup on top.

One 10-second 720p clip, 16:9, audio on, text/image references:
→ Higgsfield Cinema Studio 4.0 API list price: $4.62
→ Sume Director Studio, at the Seedance API price: $2.31

At 480p it's $2.06 vs $1.03. Director Studio runs at 480p and 720p.

No subscription either. You top up from $5 and a top-up lasts 365 days, so there are no monthly credits expiring before you use them.

The clip below is the same FPV shot side by side, with both prices on screen.

One note so the comparison stays fair: that's the Cinema Studio API vs the Seedance API price. Higgsfield's website credits are a different billing system.

If you're making AI video every day, does half the cost per clip change how many takes you'd run?

Paid partnership with Sume.

[David Im @davidim]

We just killed Higgsfield Cinema Studio.

Introducing Sume Director Studio 1.0 : the same workflow, with the original cost of seedance.

No subscription needed. Available for all users.

[Marko @Marko_Poly]

Grok Bot 1.0 says your bots now use fewer resources. An Ultra user who paid for the whole year hit 29% on day one.

Four days ago I wrote that 93.2% of one user's tokens went to bots re-reading their own chat. Then came the reset, the overnight fixes, and Matt Palmer from the team: "hard at work on token efficiency and usage." On Wednesday 1.0 shipped with this line: "We've also made them more efficient, reducing the resources required to complete tasks."

Here's what users posted the same day.

Jason Oakes, Ultra, paid for the year: "My weekly limit reset last night and I'm at 29% after the first day." His math: out by Friday, no reset until Tuesday. Two weeks ago he finished the week at 97-98%.

In the video below, his screenshot and the ones that followed:
→ Super plan: 19% in one day, "as bad as last week"
→ 90%+ used two days into the cycle
→ about 30% of one user's usage is bots talking to bots
→ an Ultra user asked for a refund, got ignored, and had Amex cancel the charge
→ Max Weinbach asked his bot about "a big update today." The reply: "usage limit reached." His meter said 0%.

To be fair, not everyone sees it. Some users say it's been better since the reset. That's the real problem with a meter nobody can see inside: two people on the same plan describe two different products.

Did your week get longer after 1.0, or shorter?

[Marko @Marko_Poly]

Elon says Grok Bot token fixes are "underway." One user already measured where his tokens went: 93.2% was the bot re-reading its own chat.

Roshel Perera runs 60 bots. On the $30 plan his weekly usage was gone in about 2.5 days, and every bot stopped for 40 hours. So he pulled one week of logs, Oct 13 to 20, and had Claude Code break it down.

In the video below:
→ 93.2% of tokens: re-reading the chat
→ 0.64%: what the bots actually said
→ 0.37%: his new prompts
→ 1,540 scheduled wake-ups a week. 45 watchdog runs, 43 found nothing

Every message and every timer makes the bot read the whole conversation again. You don't pay for the answer. You pay for the reading.

This morning the replies told the same story. An Ultra user maxes out 3 days before the reset. A Super user was at 21% a couple of hours after a reset, running one news cron.

A fix may help. But you won't know unless you measured before it lands.

Paste this into Grok Bot:
"For the next 7 days, log my usage % before and after every task and routine. Group it by bot, and flag anything that runs on a timer but finds nothing. Show me first, change nothing until I confirm."

Will the fix give you your week back, or just slow the meter down?

[Marko @Marko_Poly]

His AI sales rep costs about €38 a day. It took trial conversion from 42% to 59% in four weeks. He's still hiring salespeople.

Grok Bot 1.0 is pitched as AI teammates for a whole team. Allo already runs that way. Jérémy Goillot's phone-system startup has 21 people and about 20,000 small businesses as customers. This week he walked through the setup in an official Grok workshop.

In the video below, his own slides and his own words:
→ 26 specialist bots, each with a name, a job description, KPIs, its own phone number and email. One lead bot, "the King," manages the rest
→ support: 146 of 200 daily tickets handled by AI, 54 go to a human
→ Sarah, the AI SDR, texts, emails and now calls every trial sign-up: 42% → 59%
→ a finance bot emailed Slack's account manager and got $300 a month off. His slide puts monthly burn about $50K lower last month
→ a talent bot sourced 320 candidates in an hour. "It used to take a month"

The line people will skip: "We didn't replace our sales team." They're still hiring. The bots took the 2,000 weekly sign-ups nobody had time to call.

Two ways to read it. One founder got the output of a much bigger team without the payroll. Or every number here is his own slide, in a workshop run by the company that sells the bots.

Would you let one bot run 26 others for your company, or is that where you'd draw the line?

[Grok Bot @bot]

Introducing Grok Bot 1.0. AI teammates that work alongside your whole team.

Share Bots, collaborate with them in Slack and Teams, invite them to meetings, and watch them learn and become more helpful.

[Marko @Marko_Poly]

Someone gave Grok Bot a body for under $200. It can read every account you own and can't move a single cent.

Aleiah's "Elon Man" sits on a desk: red printed shell, a round face that blinks, a camera on top. Grok Bot is the brain. The body is an ESP32 board, three servos, a USB camera and a speakerphone, about $130 to $200 in parts. Per the guide, it's a weekend for the body and an evening for the brain.

The video below walks through the build, slide by slide.

What it does:
→ watches the office and tells you who came in
→ makes the calls you keep putting off
→ runs a money desk through Plaid: cash, bills, subscriptions, a watchlist, all read-only

The money desk is the point. In Aleiah's scan of 2,283 recent Grok Bot posts, trading and finance had the highest median views of any topic. People don't want another chatbot. They want something watching the money.

So the whole design is a gate. Send, publish, pay, delete, sign, trade, unlock: seven verbs that always come back to you as an approval card. The bot reaches the Mac only through Execution on Local Computer, set to "Ask every time."

And Elon says Grok 4.8 starts a soft rollout to Bot users this week, so the brain is about to change under it.

Paste this into Grok Bot before you build anything:
"List every action you can take on my accounts without asking me. For each one, tell me how to make it require my approval. Change nothing until I confirm."

Would you give a robot read access to your bank?

[Aleiah @AleiahLock]

https://t.co/D3BU1pIhhE

[Marko @Marko_Poly]

@bl888m_eth the funny part is when both sides are bots. two agents stuck in a phone tree together

[Marko @Marko_Poly]

The Grok Bot team's own demo call cost 7 cents. Miles Deutscher thinks you can charge $997 to set one up.

His idea: a phone-bot agency for local businesses. $997 upfront for an audit, then call bots built around the calls an owner hates most. Manufacturers, suppliers, the same client questions every day. He calls it a $300-an-hour business.

I rewatched Matt Palmer's demo from the Grok Bot team. He gave his bot a phone in about 5 minutes and had it call a coffee shop. It got the hours, confirmed they sell cappuccinos, and then hung up a little early. His words: the voice agent is "a little iffy." The bill for that call: $0.07.

The video below is that demo, with the numbers next to it.

So the math works. One audit pays for about 14,000 calls. But that's also the catch. Nobody is paying $997 for a 7-cent call. They're paying for the call that doesn't hang up early, the script that handles a phone tree, and someone to fix it when a supplier says something the bot didn't expect.

Calls are now built into Grok Bot, so setup gets easier every month. What keeps its price is knowing which calls a business makes and testing them until they stop breaking.

Paste this into Grok Bot:
"Pick a local business type I name. List the 5 phone calls it most likely makes every week, write a short script for each, and a fallback for when the call goes wrong. Don't place any real calls. Show me first, change nothing until I confirm."

Would you pay $997 for this setup, or build it yourself in 5 minutes?

[Miles Deutscher @milesdeutscher]

Matt Palmer (Grok Bot team) reveals how to use Grok Bot for live phone calls (new feature).

I actually think there's a $300/hour business idea here:

Setting up Bot call agents for local businesses.

> You conduct an initial audit ($997 upfront)
> Launch specialized call Bots
> Inject them with Skills/Routines/Routines tailored to the business

Think about it: many business owners waste a ridiculous amount of time on the phone with manufacturers, basic client FAQs, suppliers, etc.

You come in and automate as much of that process as possible.

[Marko @Marko_Poly]

@necrophosalpha honestly yeah. it got what it needed and bailed, same as me calling anywhere

[Marko @Marko_Poly]

@Shtihmas11 depends who fixes it when it breaks at 9am on a monday. that's the part you're really paying for

[Marko @Marko_Poly]

Grok Bot shipped a dozen updates this week. Four of them touch your money or your logins.

Matt Palmer from the Grok Bot team covered them in a 7.5-minute video. Here's the minute that matters.

1. It makes your phone calls. It waits in the queue, gets through the phone tree and hands you the call with a live transcript. In the demo it cancels an internet plan and turns down a 10%-off retention offer. US only, not in Washington or Illinois.

2. Checkout cards got longer and more honest. Totals now include shipping, tax and promo codes, and you get 6 hours to approve a purchase instead of 30 minutes. His reason: "I always get to the checkout and then I get some hesitation."

3. Team bots got permissions: whoever creates a shared bot decides who can change its memory. Plus a new "hours saved" chart.

4. Hidden-instruction warnings. If a page or a chat tries to slip commands to your bot, it tells you: "the page had hidden instructions for AI agents, and I didn't follow them."

The video below is those four.

Number 2 is the one I'd think about. A 6-hour window is great when you're unsure. It's also 6 hours where an approved charge is one tap away on your phone.

Paste this into Grok Bot:
"List which of this week's new features can spend my money, call someone or change a bot's memory. For each one, tell me what it can do without asking me first. Show me first, change nothing until I confirm."

Which one would you turn on first: the phone calls or the 6-hour checkout?

[matt palmer @mattyp]

What's new in Grok Bot

00:00 🌁
00:10 Phone calls
01:02 Share sheet
01:27 X cards
01:40 Shopping deals
01:58 Gifts
02:26 Team Bots
03:06 Search
03:21 Hidden-instruction warnings
03:37 Attachments on drafts
03:52 Sign-outs
04:16 Weather
05:09 Rapid fire
06:52 A note on usage

[Marko @Marko_Poly]

He took AppSumo from $3M to $84M as CEO. Now he runs about 140 Grok Bots and talks to 8 of them.

Ayman Al-Abdullah posted his whole org chart yesterday. Penny is his Chief of Staff, with bots under her for his calendar, travel, even his wardrobe. Otto runs offers, Finn is his CFO, Harvey is his general counsel. JB runs content with 13 sub-bots. Kevin runs sales. "The rest report to bots who report to me."

His line for it: "Headcount is overrated. Revenue per human is underrated." He's been saying "hiring is a sign of failure" for ten years. Now he finally has the staff to prove it.

The part I'd read twice is in his own guide. All your bots share one cloud computer. "When one bot signs into a site, every bot is signed in." And "there's no switch to stop bots messaging each other."

So the bot that picks your outfits and the bot that runs your finances sit on the same keyring. That's fine at 3 bots. At 140 it's a security model, and most people never draw it.

The video below is his chart: 8 leads, 132 bots behind them, one computer underneath.

Paste this into Grok Bot:
"List every bot I run and every site or app each one has signed into. Group them by what they can touch: money, client data, everything else. Flag any bot that can reach money but doesn't need to. Show me first, change nothing until I confirm."

Would you run a company on 140 bots, or is one keyring for all of them the first thing you'd fix?

[Ayman Al-Abdullah 🧱 @aymanalabdul]

I took AppSumo from $3M to $84M as CEO and cut my own hours in half doing it. I've been saying “hiring is a sign of failure” for 10 years.

If I'd had AI back then, I would have done $840M.

Today I run ~140 Grok Bots, and I only talk to about 8 of them. Here's how:

[Marko @Marko_Poly]

@Shtihmas11 4D chess where one of the pieces is claude

[Marko @Marko_Poly]

Elon just called Claude Opus 5.5 "a masterpiece." Your Grok Bot has been using it since October.

This morning he posted that SpaceX threw out its RL and inference stack and rewrote it in C/C++. Grok 4.8 is in RL now. Someone in the replies asked how it compares, and he answered: "Opus 5.5 is a masterpiece. But Grok 5 should exceed Opus 5.5 and maybe that happens in 2 or 3 months."

Here's the part most people scrolled past five weeks ago. On October 7 he wrote that Grok Bot "will use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs." That post did 80 million views.

So when your bot drafts a proposal or fixes a script, there's a real chance the work is done by the model Elon is trying to beat. That's fine. It's also why yesterday's "did my bot get dumber?" posts are so hard to answer. You don't know which model answered.

The video below is the router: tasks come in, one hub, five lanes out.

Theo posted a workaround last night, 84K views: tell Grok Bot to send coding work to Claude on purpose. I'd start one step earlier and just ask.

Paste this into Grok Bot:
"For every task you ran for me this week, tell me which model handled it and roughly what share of my usage it took. If you can't tell, say so. Show me first, change nothing until I confirm."

Would you wait 2-3 months for Grok 5, or route your best work to the masterpiece now?

[Elon Musk @elonmusk]

We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.

Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right. Now, 4.8 is in RL and the evals are starting to climb. We will do a soft rollout to Bot users later this week and see how it goes.

Grok 4.8 will bear no resemblance to 4.7 (thankfully). It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.

[Marko @Marko_Poly]

@hammertime_one buying from the rival while racing it. very on brand

[Marko @Marko_Poly]

@thekuchh exactly. "it got dumber" without a model name is just a vibe

[Marko @Marko_Poly]

@DhruvJain08 Elon said "a soft rollout to Bot users later this week." nothing yet on default vs opt-in. if your bots run all day, log a few tasks today so you have something to compare when it lands

[Marko @Marko_Poly]

Grok Bot just reset everyone's weekly limit. It's the 4th time since August.

August 26, when Grok Bot opened up to SuperGrok and Cursor Pro users. September 1, Elon: "another free reset on token usage." October 2. And last night: "We've reset weekly limits for all Grok Bot users. Enjoy!"

6 days, then 31, then 38. The last one came after a week of people hitting the wall by Wednesday.

I'm not complaining about a free reset. But read the four posts again and look for the one number that matters. None of them says how big the week is. Not in tokens, not in tasks, not in hours. You get a percentage, and then the percentage gets wiped.

That's why last week felt so random. Roshel Perera ran out on the $30 plan in about 2.5 days. Someone else was at 19% after a weekend without coding. Nobody could compare, because nobody knew what 100% was.

The video below is the loop: the meter fills, the post goes out, back to zero, four times.

The team says it's sprinting this week to make usage "go farther." Good. The only way you'll know if it worked is your own log, started before the next reset wipes it.

Paste this into Grok Bot:
"Starting now, keep a daily log: my usage % at 9am and 9pm, and the 3 tasks that used the most. Don't change anything. On Sunday, send me one table: what my week actually bought, in tasks. Show me first, change nothing until I confirm."

Is a reset every few weeks a gift, or a sign the meter still isn't working?

[Grok Bot @bot]

We've reset weekly limits for all Grok Bot users.

Enjoy!

[Marko @Marko_Poly]

A real estate business now runs on 4 people instead of 50

Cole Ruud Johnson posted his setup last night. He talks to one Grok Bot, a Chief of Staff. It manages a bot for each department, and each department has 3 to 5 bots under it. CRM runs the pipeline and reviews sales calls. Marketing handles the website, reviews and SEO. Relationships reaches out to attorneys, lenders and agents for deal flow. His words: "more profitable than ever."

I believe the org chart. What I'd want to see is the line under it.

Every one of those bots draws from one weekly limit. Last week Roshel Perera's 60 bots burned through his in about 2.5 days, and all of them stopped for 40 hours. Nothing crashed. His whole operation just went quiet at once.

With 50 people, a bad week means a few people are out. With 4 people and an org chart of bots, it means every department stops on the same afternoon.

The video below shows it: 46 seats go dark, the bots take over, then the meter hits 100%.

That's not a reason not to build it. It's a reason to know who picks up the phone when it happens.

Paste this into Grok Bot:
"List every bot I run and what it does. Mark which tasks touch clients or money. If my weekly limit hit 100% today, tell me which of those would stop, and draft a manual fallback for each one: who does it and how, in under 10 minutes. Show me first, change nothing until I confirm."

Is 50 to 4 the future of small business, or a company one meter away from zero?

[Cole Ruud-Johnson @coleruudjohnson]

I think what I've set up in Grok Bot for my real estate business is peak AI.

We've reduced our headcount from 50 people to 4, and are more profitable than ever.

I talk to one bot (Chief of Staff), and it manages a bot that runs each department, and each department has 3-5 bots.

Some of my favorites:

-- CRM Bot: Manages pipelines, CRM, team activity, sales call reviews, pipeline reviews, KPIs, and making sure no opportunity goes missed.

-- Marketing: Handles our website, reviews, brand, testimonials, landing pages, SEO ranking, credibility packets, and more.

-- Relationships: Reaches out to attorneys, lenders, agents, & other investors to create free deal flow.

There are many more..

Comment "COS" below and I'll DM you links to download & copy my bots + our whole bot org chart for free.

[Marko @Marko_Poly]

Elon said the Grok Bot fixes would go in overnight, with a "dramatic improvement by tomorrow." The meters moved. Whether the bots got better is a different question.

What people posted after the fix:
→ one user reset on Sunday, hit 26% a few hours later, and woke up at 2%
→ another went from 27% in the morning to 1% by evening
→ Steven Mark Ryan: "MASSIVE reduction in token usage per task"

That's the good half. The other half is in the same replies. One user says it kept answering "I don't know" until they pushed it to dig deeper. Another says part of their usage still goes to "fixing its own mistakes." One ran a single small task and says it got worse.

Last night the team said they were working on "usage limits and model routing." Routing is the part I'd watch. A task can get cheaper because the bot reads less junk, or because it quietly goes to a smaller model. The meter can't tell you which.

The video below is today on one screen: the meter, cost per task, and the complaints in the order they came in.

So don't judge the fix by the percentage. Judge it by the work.

Paste this into Grok Bot:
"Take 3 tasks you did for me last week. Run them again today the same way. For each one, show my usage % before and after, and put last week's result next to today's. Tell me honestly which one is worse. Show me first, change nothing until I confirm."

Did your bots get cheaper, or just lazier?

[matt palmer @mattyp]

Improvements to token efficiency & behavior in Grok Bot are underway

Many of you have built important workflows on @bot - we want to keep supporting you

Please reply with feedback on this post. I will read it all

Thank you sincerely for your support

[Marko @Marko_Poly]

@ChewieWins over a quarter by monday with a friday renewal is rough. that's the week a log earns its keep

[Marko @Marko_Poly]

Grok Bot users are waking up to a reset: weekly usage back to 0%. Before you burn it again, give one bot a single job: watch the others.

Yesterday's numbers made the case. One user pulled a week of logs from 60 bots: 93.2% of tokens went to re-reading chats, 0.64% to what the bots actually said. 1,540 timer wake-ups. 45 watchdog runs, 43 found nothing. All from one shared weekly limit, with no breakdown in the app.

And the reset doesn't change the habit. One SuperGrok Heavy user was back at 14% within an hour this morning, just tidying up his bots.

That's what this bot is for. In the video below:
→ it logs your usage % before and after every task, per bot
→ it flags timers that wake up, read everything and find nothing
→ at 70% it pauses the nice-to-have bots
→ at 85% client work goes to you, not to a bot that's about to run dry
→ every morning you get one report: who burned what

A reset gives you the week back once. It doesn't tell you where the last one went. This does.

Paste this into Grok Bot:
"Create a new bot called Warden. Its only job is to watch my other bots. Log my usage % before and after every task and routine, grouped by bot. Flag any timer that ran but found nothing. At 70% weekly usage, pause bots I mark as nice-to-have. At 85%, stop client-facing work and message me instead. Send me a one-screen report every morning at 8. Show me first, change nothing until I confirm."

Would you trust one bot to switch off the others?

[Marko @Marko_Poly]

Elon says Grok Bot token fixes are "underway." One user already measured where his tokens went: 93.2% was the bot re-reading its own chat.

Roshel Perera runs 60 bots. On the $30 plan his weekly usage was gone in about 2.5 days, and every bot stopped for 40 hours. So he pulled one week of logs, Oct 13 to 20, and had Claude Code break it down.

In the video below:
→ 93.2% of tokens: re-reading the chat
→ 0.64%: what the bots actually said
→ 0.37%: his new prompts
→ 1,540 scheduled wake-ups a week. 45 watchdog runs, 43 found nothing

Every message and every timer makes the bot read the whole conversation again. You don't pay for the answer. You pay for the reading.

This morning the replies told the same story. An Ultra user maxes out 3 days before the reset. A Super user was at 21% a couple of hours after a reset, running one news cron.

A fix may help. But you won't know unless you measured before it lands.

Paste this into Grok Bot:
"For the next 7 days, log my usage % before and after every task and routine. Group it by bot, and flag anything that runs on a timer but finds nothing. Show me first, change nothing until I confirm."

Will the fix give you your week back, or just slow the meter down?