If your Grok Bot @bot usage is insufficient, don't rush to upgrade. Let's break it down by mechanism: how to bill, who burns the most, how to audit, and how to prevent leaks. It's not based on the number of messages you send. The official description is: it's based on agent steps and tokens; the more activity, the more expensive. High-frequency routines, long conversations with rereading each round, browser clicks, and multi-bot group chat replies—these are where the real credit limit is consumed. Official statement: Hourly routines can consume a week's credit limit in a single day. Let's break it down by mechanism: how to bill, who burns the most, how to audit, and how to prevent leaks. 1/ First, distinguish between the two ledgers; don't mix them. X Subscription (Premium / Super / Ultra): Cursor + Grok Bot + Grok apps share a weekly credit limit, cleared weekly; unused limits are not carried over. Once used up, it uses prepaid additional usage. Another option for the old plan: Weekly usage limit + On-demand, billed in your Cursor account. SuperGrok/Premium+ and Cursor plans do not stack; use the larger plan. Common point: Deductions are made based on steps/tokens, not message count. Official website: http://((cursor.com))/help/grok-bot/plans 2/ Which consumes the most resources, ranked by mechanism (not percentages): ① Routines: Deducted every time they run, even if they are idle. Pausing/never running = no deduction. Clicking "Test" also counts as actual work. ② Long conversations reread every round: Forum staff confirmed: The longer the history, the more expensive each message; idle old chats not sent do not incur deductions. In the CSV exported by the community, cache reads are common—this is a user sample, not an official percentage. ③ Browser/desktop screenshot loops: Officially clarified: Connectors are used first if available. ④ Group Chat + Bot ↔ Bot Every Bot reply counts, including replies in the group chat and messages exchanged between Bots. The fewer Bots, the fewer replies. ⑤ Wide hints, frequent retries, raw images and videos in Bots There is no "Choose Cheap Model" switch in the Bot. Raw images under X subscriptions may still be used in the same weekly pool. 3/ Don't guess, measure first. ① Pause unnecessary routines (don't rush to delete them), see if the weekly usage percentage changes. ② X subscription: http://(grok.com)/dashboard → Usage; Old plan: Settings → Usage & Billing ③ Export to CSV if possible, and ask the main Bot: Which conversation, which routine, which day was the most expensive? ④ Don't create a "check usage every N minutes" routine—checking itself consumes tokens. The exact weekly token usage for each tier is not officially disclosed. Don't trust the absolute numbers online. 4/ Trading Checklist. Official order: First pause / lengthen the interval / narrow the trigger, then discuss upgrading. Pause takes precedence over deletion (recoverable, no deduction during pause) 15-minute/30-minute monitoring → Change to hourly, or run 2-3 times on weekdays Avoid indiscriminate polling if event-triggered (Slack keywords, GitHub, webhooks) Narrowly define triggers: "deploys include FAILED", avoid "any message in busy channel" One task, one conversation; write persistent rules in the Bot description or Skill, don't pile them in chat Use connectors if available, use browsers as a last resort Minimize @ mentions and Bot chat in groups; change roles that don't require independent routines to Skills Write results to a file, read the file next time, don't let it reread dozens of pages of chat every round Auto Top-up/On-demand: Set your own monthly limit Official Routine Guide: http://((cursor.com))/help/grok-bot/routines 5/ New project launches/project scanning are particularly prone to missing: One-time, on-time (mint open) > All-day polling Scanning: Only push if there are changes, don't 15-Minute Full Study If you can read via API/connector, don't open a browser to click on the page. Don't wake up a second bot just to "confirm." Write the log file; read the file in the next round. "Don't speak if there's no change" only shortens the response time, but still incurs the wake-up cost. This is a community consensus and aligns with the official "deduct every time it runs" policy. 6/ Anti-patterns (official + community aligned pitfalls): * 24/24 "casual browsing" routines * Slack "any message" wide triggering * One chat for everything, the more you chat, the more expensive it gets * A bunch of bots in the same group tagging each other * Having a connector but still opening a browser to take screenshots * Frequently clicking Test Routines * Thinking that sending it to the Cursor Agent will avoid deducting from the weekly pool (X subscriptions are in the same pool) * Hoping to choose a cheaper model in the bot—there's no such switch * Using a usage-checking routine to save usage—putting the cart before the horse 7/ Insufficient usage is often a clock problem, not a stingy plan. Fix the leaks first, then talk about Ultra. Upgrading doesn't fix wake-up calls, it just increases leakage. Risks are also clearly stated: overnight submissions are considered drafts; actions like sending money, deleting, or outsourcing should not be always allowed; don't force approvals after they expire. The above is compiled based on official help and forum staff instructions. Absolute limits and percentage savings are not listed for anything not publicly disclosed.