Grok Bot’s New Weekly Pool, Voice Upgrade, and Desktop Warnings

James Turner

Welcome back to Grok Bot Release Notes, brought to you by Jellypod AI. Last time we covered version 0.70.2, so I won't rehash the bot inboxes and unlinked X search. Today is about what landed between October 12 and October 16. The biggest change isn't a feature in the app. It's how you pay for it.

James Turner

On October 12, SpaceXSI and Cursor announced The X Subscription. It folds Grok, Grok Bot, Cursor, and X Premium into one plan. The tiers are Free at zero dollars a month, Plus at eight, Premium at thirty, Super at one hundred, and Ultra at two hundred. Premium carries four times the usage of Plus. Super carries fifteen times, and Ultra fifty times. Teams get Premium, Super, and Ultra at the same per seat prices, thirty, one hundred, and two hundred. That's described as cheaper than legacy Cursor Teams, with three times more coding usage.

James Turner

And that Premium price replaces something specific. SuperGrok, Cursor Pro, and X Premium Plus used to cost ninety dollars a month combined. Now it's thirty.

James Turner

Here's the mechanism that matters for Grok Bot. Paid plans now share one Grok model usage pool across Grok Bot, Cursor, and the Grok tab on X. And the allowance refreshes weekly, not monthly. In the official announcement, the line is that one heavy day doesn't use up your month. Think about a routine that goes wrong overnight. Under a monthly budget, it could drain the whole team's month. Under a weekly pool, the damage resets in days.

James Turner

Now the trade off, and this is where the community is pushing back. Third party models are no longer included in the flat pool on the new X plans. You pay for them at API list prices through top ups. My interpretation, and it's only my read, is that this nudges coding and bot workflows toward Grok 4.7, since that's what the pool actually covers.

James Turner

On the Cursor forum, the announcement thread has well over a hundred replies. Users named Archit, coco_max, and tangjun are among those reporting early confusion. The common complaint is that heavy Grok Bot workflows burn through the shared weekly quota faster than expected. One feature request on the forum asks for per agent usage, meaning who worked, tokens, cache, and cost against the weekly pool. It had six votes when we checked. Another thread, from an Ultra user, says their weekly allowance hit ninety nine percent in three days.

James Turner

So here's how I'd use this. If you run several bots, treat the weekly pool like a budget. Put your most expensive routines on a schedule you can see. Check usage midweek, not at the end. And remember Grok Bot's own Usage and Billing screen now shows your plan name and labels its percentage as used, which came in with 0.70.2.

James Turner

Second change: on October 14, xAI introduced Grok Voice 3.0, and it also powers Grok Bot's voice mode. So when you call your bot, you're now talking to the new model. xAI bills it as the most intelligent speech to speech model, with latency under a second. On Artificial Analysis's Speech to Speech Index it scores 86.9, against 81.3 for Voice 2.0. On Full Duplex Bench, which tests interruptions and turn taking, it scored 97.0 percent.

James Turner

What does that mean for you? Interrupting your bot mid sentence should feel less awkward. For developers, voice agent calls upgrade automatically under the model name grok voice latest, at the existing rate of eight cents a minute, which is four dollars eighty an hour.

James Turner

SpaceXSI also gave real world numbers. Voice 3.0 handles millions of daily driver requests in Tesla vehicles, things like calendar checks, emails, coffee orders, and reservations through Grok Bot. And in Starlink call centers, it answers twenty thousand calls a day, with about seventy percent resolved without a human. Those are SpaceXSI's own figures, so read them as company claims.

James Turner

Now the desktop client. Version 0.71.2, built October 14, has been rolling out to macOS and Windows, and it arrived without an official changelog entry. As of our check, the official log still tops out at 0.70.2. So everything I'm about to say about 0.71.2 comes from forum reports, not release notes.

James Turner

The biggest gotcha is on Windows 11. A user named ratiphi reports that local tool access is blocked, things like the CopyFromBox action and local shell execution. The error reads, Local tools are turned off, the user has set local tool access to Never. But in Settings, under Execution on Local Computer, it's set to Always Allow. A separate thread describes the same stuck on Never behavior on a Mac, for one account only. The workaround for now is to fall back to in chat attachments, or run the work on the cloud VM until a patch lands.

James Turner

Two smaller reports. Newly created bots in 0.71.2 default to Unassigned with no category dropdown, and the avatar editor locks to a default gray shape with no custom shape or accent color pickers.

James Turner

Then there's the more serious category, the cloud computer. Some users describe their Agent Computer becoming unreachable, with a missing gateway descriptor file, leaving bots in an unresponsive loop where manual resets fail. And several users report silent rollbacks. One user said their cloud computer reverted twice on October 8, losing about two hours of video rendering. Another thread claims a full disk triggers a silent rollback and loses two days of work. A support engineer hasn't confirmed that cause, so treat it as a user hypothesis. The practical lesson doesn't need confirmation though. Commit code to an external git repository, or sync to a cloud drive, continuously. Don't let the VM be the only copy.

James Turner

On the positive side, creator Matt Palmer's October 13 walkthrough highlights a Shopify connector. His description is that your bot can track orders, check your store's status, and read your metrics. That's from his video, not the official log, so check your own connector list before counting on it.

James Turner

Last, two safeguards from the 0.70.2 notes that matter more now that bots can pay with X Money or Stripe Link. Before an action it can't undo, like sending, posting, buying, or deleting, a bot checks your own messages for approval. And after a bank or brokerage connector returns data, the bot asks before sharing it anywhere else. Also, the notes say Grok Bot can't reach a private Tailscale MCP server yet, and suggest Tailscale Funnel to expose it.

James Turner

So, the takeaways. Watch your weekly pool. Try the new voice mode. If you're on Windows 11 and 0.71.2, expect local tool trouble. And back up your work outside the VM. That's it for today, and this episode was brought to you by Jellypod AI.