Translated from Portuguese by AI Original

Grok Bot the Right Way


0
This article was translated to English using AI.

How much does an employee who wakes up 96 times a day, opens the fridge, finds nothing new and goes back to sleep cost? If they’re a bot, the answer shows up on your subscription meter, and on the Halloween weekend a lot of people discovered the price in the most unpleasant way possible: an entire week of usage evaporated in a few days .

A bot opening the fridge at night — just like me.

Users reported meters jumping from 50% to 100% overnight, and Musk himself commented that serving the Grok Bot turned out to be too expensive relative to revenue. No reset was promised (at least not without you handing over your data).

There was also, according to a tracker’s report, a billing dashboard bug on the same day, so part of the scare may have been a display error. What didn’t change is that the usage pool isn’t going to get bigger anytime soon.

I use Grok Bot with 53 automated routines spread across 17 bots and I’ve never hit 100% of the quota. I don’t use it for development, just for the rest of life. I even have an orchestrator bot that every week suggests optimizations for the other routines (yes, a manager of managers, every corporate dream). This piece is what I learned trying to keep this mess cheap, with the help of an independent study paper whose claims I checked against the official documentation (what I couldn’t confirm is left out).


One pool, one number, no clues

Under The X Subscription, the Grok Bot uses the same weekly pool as Cursor and the Grok apps , with no separate meter. You see one percentage and the reset date. Grok Bot is included in the Premium, Super and Ultra plans, and excluded from Plus.

The entry package that includes Grok Bot is Premium, which costs US$ 30 per month and bundles the benefits of SuperGrok, Cursor Pro and X Premium+ (which together used to cost US$ 90 per month), plus the verified badge on X. Its usage pool is 4× the Plus. The tiers above are Super and Ultra:

PlanPrice per monthUsage pool
PlusUS$ 81x (no Grok Bot)
PremiumUS$ 304x
SuperUS$ 10015x
UltraUS$ 20050x

The package was announced on October 12, 2026 . Existing SuperGrok, Cursor or X Premium subscribers can migrate via billing settings, with proportional refunds for the old plan, or simply keep what they have. For teams, Premium, Super and Ultra are charged the same per-seat price as individual plans.

The meter lives at grok.com/dashboard, under the Usage tab. (If you entered via the Cursor plan, the path is different: Grok Bot, Settings, Usage & Billing.) Until recently it only showed the total, without saying which bot used what. That changed: the Usage tab has started showing a per-bot breakdown (it arrived a few days ago and it’s nice to see that the bot using the browser burns many tokens). You can now point to the greedy bot, but it doesn’t say what inside it is heavy.

And guesses tend to be wrong, because the natural reaction is to cut what you see: ask the bot fewer questions. Meanwhile, the routine that wakes every 15 minutes to say “nothing new” stays there, working in silence with the dedication of an intern without a boss.


The bill nobody shows

The documentation doesn’t explain how usage is calculated, so what follows is a mental model, not a billing formula. It helps decide where to cut, and it works like this: the cost of a task is roughly the result of multiplying four things.

  1. How many times the bot wakes up (each routine, test or message).
  2. How much it needs to reread at each step (profile, memory, old conversation).
  3. How many steps it takes each time (clicks, attempts, long response).
  4. How many bots answer the same question.

A bot showing a dashboard nobody understands.

The important word is multiply. A routine that wakes 672 times a week, inside a six-week conversation, doesn’t cost a bit more than a lean routine. It costs several times more. The good news is the reverse also holds: halving wakeups, cutting what the bot rereads to a third, steps to a third and duplicate bots to 10% can drag a full pool down near 20%. Four modest savings, none heroic.

Each tip below tweaks one of those four factors.


Before you cut, ask the bot itself

A bot pondering a girl’s question.

The Usage tab already shows spend per bot, but not what inside each one is heavy. For that, the bot itself can list what it has, and that’s enough to rank suspects. (My orchestrator runs a version of this every week.) Paste into your main bot:

Quero entender onde está indo o meu uso semanal. Sem mudar nada, me traga:
1. Todas as rotinas que você e meus outros bots têm: bot, nome,
   agenda ou trigger, e quantas vezes rodam por semana.
2. Os bots com a conversa mais antiga e mais longa.
3. Tarefas que você faz abrindo site e que poderiam ser feitas por
   um app conectado.
4. Grupos em que mais de um bot responde à mesma pergunta.
5. Um ranking com os 5 itens que mais devem gastar, com uma linha
   explicando cada um.
Espere o meu OK antes de alterar qualquer coisa.

Tip 1: stop waking the bot unnecessarily

A routine is a task the bot runs on its own, at a schedule or when something happens. The documentation is blunt: each run consumes usage, and examples of high spend are hourly calendars, short intervals and Slack triggers in busy channels. The weekly math is scary:

ScheduleExecutions per week
Every 15 minutes672
Every 30 minutes336
Hourly168
Every 2 hours84
3 times per workday15

Three adjustments solve most of it.

Slow the schedule down. Nothing should run more often than hourly unless you have a good reason. Work routines only on weekdays, during your hours. For that the timezone must be correct, and it’s under Settings, in the Bot section, in the Timezone field. To pause a routine, open the bot, go to the Details tab and turn off the switch on the line (on mobile, turn off Active).

Swap schedules for triggers. A schedule wakes the bot even when nothing changed. A trigger wakes only when something happens. Available triggers are Slack messages, GitHub, Linear, Sentry events (direct use — wonderful!), PagerDuty, email to the bot’s address and webhooks. In Slack, pick the narrowest trigger, like “I was mentioned” or a specific phrase, and avoid “any message” (the digital equivalent of arming your house alarm to go off when a leaf falls in the yard).

Don’t test in production. The Test button runs the routine for real, consumes usage and, according to the documentation, can browse sites, change files and call connected tools. Write and adjust the instruction in chat, and test only once.

Liste as suas rotinas e as dos meus outros bots com a frequência
atual. Depois proponha a agenda mais lenta que ainda resolve cada
tarefa:
- nada mais frequente que de hora em hora sem o meu OK;
- rotinas de trabalho só de segunda a sexta, no meu horário comercial;
- junte rotinas que leem a mesma fonte;
- apague as de "avisar quando X acontecer" que já aconteceram;
- se a rotina só espera um evento, sugira trocar a agenda por um
  trigger e diga qual.
Mostre uma tabela de antes e depois com execuções por semana. Não
mude nada até eu aprovar.

Tip 2: old bots reread too much

Each reply causes the model to reread the conversation. Much of that is cached, which is cheaper, but cheaper isn’t free. A bot that’s been doing the same job for six weeks carries six weeks of conversation into every reply, so the same task that was cheap the first week becomes noticeably heavier by week six, even if nothing changed.

A bot reading in a beautiful reading room

There’s no manual compact button for conversations. There are two exits, both described in the docs :

  • Duplicate the bot (the page mentions the iPhone) copies profile, settings, skills and routines, but doesn’t copy conversation history or memory. (I’ve never used it and never will, despite people saying it’s good practice — I’ve grown attached to my bots)
  • Reset the conversation clears history but keeps memory, files, routines and settings. (this one is worth it)

Before either, ask for a handover note. If you reset, also ask the bot to save the rules that matter into memory.

Escreva uma nota de passagem para um bot novo assumir o trabalho
"[nome do trabalho]". Inclua:
- o objetivo e como saber que terminou;
- as minhas regras e preferências para esse trabalho;
- os apps e contas usados;
- o estado atual e as pendências;
- os erros que já cometemos.
No máximo 300 palavras, sem histórico de conversa e sem conversa fiada.

Two complementary cleanups. First is memory: there’s no screen for it, so you chat with the bot, ask for a list of what it stored and tell it to correct or delete what’s stale. Second is profile: personal bots don’t have a settings page anymore, so you change behavior by telling the bot, in chat, to follow short rules.

And when one bot does inbox, reporting and research at the same time, it carries the context of all three into every step. One job per bot keeps each conversation short. The savings only show if bots don’t spend the day chatting with each other, and that’s tip 4.


Tip 3: a task with a finish line

A bot reaching its destination: the summit of a mountain

“Take a look at this” is the most expensive phrase there is, because it has no end. The bot keeps exploring, trying again, opening another tab. A task template solves it:

Tarefa: [uma frase]
Escopo: somente [arquivos / canal / caixa de entrada / lista].
Pronto quando: [critério que dá para conferir, como "tabela de 10 linhas"].
Pare se: o mesmo passo falhar duas vezes, ou faltar algo que eu não
te passei. Nesse caso, me diga o que está travando.
Não faça: pesquisar alternativas, abrir abas extras ou chamar bots
ajudantes.
Resposta: [formato], no máximo [N] palavras.

(Anyone who’s worked with Scrum will recognize “Pronto quando” as the Definition of Done in other clothes. Agility hiding inside a prompt.)

To make it apply across tasks, save one stop-rule into memory once:

Salve na sua memória, para todas as tarefas: se o mesmo passo falhar
duas vezes, pare e me avise; se faltar informação, pergunte uma vez
só; em tarefas de menos de uma hora, mande apenas o resultado final,
sem mensagens de progresso.

If a bot enters a loop, send a direct message with “Stop now”. The documentation says a message from you takes priority over background work and ends the job immediately (without undoing what it already completed).

Three reinforcements help keep steps low:

  • Apps instead of clicks. From the bot’s computer, the path is open the site, log in, look at the screen, find the menu, click and look again. With a connected app, the same thing becomes a direct call, and the documentation says the app is more reliable than clicking a site when one exists. You connect under Connect apps, authorize login, and you can even ask “add the Linear app” in chat. A connected app applies to all your bots.
  • Skills for repeated work. If you explain the same report every Monday, the bot redoes the reasoning every Monday. Ask it to save the process as a skill, in short steps, and call it by name or with @. There’s also Teach a task, which records up to ten minutes of your flow to become a skill, but rollout is gradual.
  • Short responses. Ask for a fixed format and a word limit. A 600-word report when you read the first three lines is waste without return.

Tip 4: one owner per conversation

Every bot reply counts toward your usage, including in groups and including messages a bot sends to another bot. A group has 2 to 6 bots, so a question in a group of four can generate four replies. To add, there’s no button to stop bots from talking to bots. The documentation suggests asking for that in chat.

ScenarioBots in chatResponses
Group, everyone replies4up to 4
Group, only the owner replies41
Direct message to the owner11

So put the rule in writing, and have the bot save it:

Novas regras para o trabalho em equipe dos meus bots. Salve na memória
e passe aos outros bots.
1. Um dono por trabalho. Em grupo, só o dono responde; os outros ficam
   quietos, a menos que eu marque com @.
2. Passe o trabalho uma vez, já pronto. Sem atualizações de progresso
   entre vocês.
3. No máximo 2 mensagens entre bots por tarefa. Se precisar de mais,
   me pergunte.
4. Sem bots ajudantes para o que você consegue fazer sozinho.
5. Se não tem nada novo a dizer, não responda.
Confirme cada regra em uma linha.

Keep only the bots in groups that the work requires (two usually suffice) and send ad-hoc questions directly to a single bot, not through the group.


All together: from 100% to 20% (in an imaginary world)

A bot lying in a hammock, relaxed and sipping a drink — with a beach in the background

The paper I used as a base constructs an example to show how the tips add up. A small company with four bots, using invented “units” (the week before changes equals 2,000 units, or 100% of the pool). The numbers are illustrative. No one publishes the real cost of a run.

WorkWhat changedBeforeAfter
Inbox watcher15-min schedule became email trigger67215
Competitor monitorFrom every 30 min to 3 times per workday33615
Daily reportNew bot with handover note, skill and app22442
Team groupOnly the owner replies32080
Ad-hoc tasksFinish line and stop-rule360240
Browser jobsBecame app calls888
Total2,000400

Two observations. First: alone, tip 1 dropped the 100% week to about 51%, because two routines that almost never found anything consumed half the pool. It was two configuration changes and no loss of function. Second: the work the person actually asked for stayed and went from a third of the use to nearly three-quarters of what remained. The week now mainly goes to things someone wanted.

A common-sense caveat: a setup without many routines hardly gains from tip 1 and benefits more from tips 2 and 3. What matters for any setup is order. Audit first and cut the factor the ranking shows is biggest.


What no one knows yet

SpaceXAI doesn’t publish pool size, the cost per execution, or which model the bot used. The plan docs also don’t mention warning before hitting 100%: the only message described appears when the limit is reached. The November 1 reports are user accounts and weren’t verified, and some features, like Teach a task, are still rolling out, so configuration paths may change.

Each tip also has its price. A slower schedule delays things. A new bot loses some nuance that the handover note didn’t capture. A stop-rule sometimes cuts a task that a third try would have solved. A single owner loses a second opinion. Keep the fast pace or the extra bot where the work really needs it, and pay for that on purpose.

Whoever exceeds the pool has a decision to make. Under The X Subscription, extra usage comes out of prepaid credits, and you can set a monthly limit at grok.com/dashboard/billing, in Additional usage credits. Leave that limit set and Auto Top-up off unless the bot needs to run at any cost.


Who audits the auditor

A bigger bot auditing a smaller bot

The pool won’t grow by itself, so the way is to slim the bots. Almost all waste is invisible because it doesn’t look like work: a routine waking up to find nothing, a bot rereading six weeks of conversation, four bots answering the same question, a bot clicking a site that an app would read with a single call.

That leaves the question no paper answered: who audits the bot that audits the bots? For now it’s me, and I also open the fridge sometimes just to check what’s inside.