Grok Bot: SpaceXAI's AI Coworkers Get Their Own Computers

On November 3, 2026, G​rok Bot stopped being a product that only ever talks to you. “Try asking G​rok @Bot to make calls!” Elon Musk posted that day, and what he was pointing at is narrower than a phone assistant and stranger than it sounds: once you approve a call, the Bot dials a business, waits out the hold queue, presses its way through the automated menu, talks to whoever picks up, and hands you a summary of what was said. The agent doing that is the same G​rok Bot whose reasoning has moved from Grok 4.5 to Grok 4.7 since launch — and which, on its owner's account since October 7, may hand the hardest work to Claude Opus 5.5 — but a phone line is the first place it acts where none of your logins apply.

The capability was announced on October 25 and spent nine days in preview builds. What moved on November 3 is who can reach it: creator Peter Yang said that day that all G​rok Bot users should now be able to call US businesses, and Musk spent the day telling people to try it. The footprint is US numbers only, and both Washington and Illinois are excluded — on the user side and on the number side. Every account of the rollout that names the exclusions names the same two states; none of them says why.

Almost nothing about the feature is documented by the company that ships it. There is no changelog entry for calling, no post on the news feed, and no page in the documentation. What exists is a company owner's posts, a set of user demos, and third-party coverage of both — the same shape as the last three features this product shipped, and the reason every claim below is labelled for where it came from.

The product behind that capability went into early beta on August 11, 2026: an AI agent product built on the idea that an AI should not just answer you, but take the work off your plate. Each Bot gets its own dedicated computer in the cloud, signs into the Gmail, CRM, and web apps your team already uses, and completes jobs end to end — coming back only when a decision or approval is needed. Ten days later, on August 21, the beta widened from the three most expensive premium tiers to SuperGrok Plus, Cursor Pro+, and every Cursor Teams plan, with a limited free trial for everyone else. Then, on August 26, S​paceXAI removed the tier gate altogether: the company announced on X that every SuperGrok and Cursor Pro subscriber — the $30/month base tier and the $20/month Cursor Pro tier included — now has G​rok Bot access, with weekly usage limits reset for all users. A week later, on September 2, S​paceXAI brought G​rok Bot to Android, listing the agent in the Play Store. On September 17 the product gained a voice: voice calls with a single Bot, on desktop first and then on iOS and Android over the days that followed. Since then the product has kept moving under the same subscription rather than behind a new price: Team Bots, shared Bots a whole team can work with, shipped on September 28, and in the week since, users have started meeting a new unpaid-for role called the Primary Bot — which the G​rok Bot account on X described on October 1 as a Bot that spots work and offers to handle it, in the same week S​paceXAI reset every user's weekly usage limits. On October 7 the product's reach changed again: G​rok Bot gained native search, read and monitoring access to X, so a single prompt can gather what the platform is saying without the account connector the August build required. Two days later, on October 9, the account that posts the product's own updates announced something that sounds smaller and reaches further: a Bot can claim an email address of its own, ending in @mail.grokbot.com, which it can use to “sign up for services, contact businesses for you, or schedule time with someone.” You claim it by asking the Bot in chat or tagging @bot on X, and the announcement described it as “rolling out to users today” — which is the whole of what is known: 9to5Mac and TeslaNorth carried it the same day, the rollout is described as partial rather than complete, and the documentation had no inbox page as of October 10, though the October 11 changelog now describes the capability itself. Three days after that the product's commercial shape changed rather than its feature list: on October 12 S​paceXAI announced The X Subscription, folding G​rok, G​rok Bot, Cursor and X Premium into one plan — and the plan pages answer, for the first time, how a Bot's work is metered. The browser had already become a fourth place to use the product: the vendor's own release notes put G​rok Bot on grok.com for SuperGrok Heavy on September 19 and widened it to SuperGrok Plus on October 3, with Bots in the sidebar's Bots section. What is not established is the other surface a third-party account claimed on October 16: Bots inside the G​rok mobile app itself, which nothing S​paceXAI publishes describes. The engine has moved too — Grok 4.5 was the model at launch, and the line has since reached Grok 4.7 — and on October 7 that stopped being exclusive. In a post on X rather than a product announcement, Elon Musk said the company would “use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs,” because “whatever is most likely to give you the best outcome” is what the Bot should reach for — A​nthropic's Claude Opus 5.5 for the heavy reasoning, Midjourney for images and Suno for music. The statement is now corroborated as a statement — The Information, Gizmodo, 9to5Mac, Android Headlines, TechRepublic and the music-industry trade press all carried it inside two days — and it is still only a statement: no rate card, no changelog entry, no shipped artefact, no outside measurement. It also does not settle what is out of scope. The named list is A​nthropic's Claude Opus 5.5, Midjourney and Suno, with everything else left as “other leading APIs”; nothing in the post excludes any other lab's models and nothing in it includes them, and no outlet has reported either way, so treat a lab being excluded as an assumption rather than a fact. What is not in doubt is the direction: the thing sitting under G​rok Bot may no longer be a S​paceXAI model. That combination — a product with a computer, and a frontier model doing the thinking — is why this launch is more interesting than another model release. What nobody outside S​paceXAI has a measurement of is how the two halves behave together when the frontier model is somebody else's — and the first real evidence about the computer half arrived on October 6, in a Slack channel. On October 21 the deliverable changed: the product's own account announced that G​rok Bot “can now build reports, memos, and dashboards. Ask Bot for a page, then edit it together or have Bot keep it up to date.” Vendor-stated, one post and a video, and the first time that account has sold the deliverable, rather than the agent, as the thing that stays current — though the changelog caught up the next day, on October 22, with a line that repeats the claim almost word for word. Six days later, on October 27, the product moved on what a Bot knows rather than what it produces: a control spotted in the desktop app by TestingCatalog offers to import another assistant's memory — “Transfer your preferences and context from another AI assistant” — and the changelog of October 26 now gives it one line.

It is also the first product to come out of the x​AI–Cursor merger. SpaceX agreed in June 2026 to acquire Anysphere, the maker of Cursor, for $60 billion in stock; Grok 4.5 was the first model built with Cursor's developer data, and G​rok Bot is the first co-branded product from the combined S​paceXAI. What began as an internal prototype — sales, marketing, office ops, bug fixing — is now the company's public bet on how AI agents should be built. The bet is specific: give each agent a computer, let it use your tools the way a person does, and let it keep working when you close your laptop.

What G​rok Bot actually is

"AI teammates you can give real work to" is S​paceXAI's own description, and the mechanics matter more than the slogan. Every Bot runs on its own cloud computer, which means it can operate software directly inside the tools where the job actually happens — navigating a website, entering information, working a portal — through the same interfaces a human would use. That is the architectural difference from most agents, which need a clean API or an MCP connector to exist before they can touch a given service: G​rok Bot needs the web UI to exist, and that is it. Where a connector does exist the documentation steers Bots toward it — they “use connectors where available and computer use for everything else” — and X is where the second path arrived, in two steps: a connector on August 29, then a native route on October 7. The inbox announced on October 9 is the same logic applied to identity rather than to access, and it is the sharper break with the launch framing. A Bot signing into your Gmail is working as you, through credentials you own and can revoke. A Bot holding an address at @mail.grokbot.com is a party in its own right: it can register for a service, write to a business, or hold a thread you were never on, which is exactly what the product is sold to do and exactly why the accounts it creates are invisible to anyone auditing your stack. It is also the cleanest example yet of the product outrunning its own manual — the G​rok Bot documentation still has no page about the inbox itself — the closest it comes is an Agent Email toggle in the Teams and Enterprise reference, off by default, which lets members claim an address and lets their Bots read and send mail from it — eleven days after the October 11 changelog entry described the capability. The launch post puts it plainly: "They sign in to your tools, use them just like you do, and come back with finished work."

What a Bot hands back has been the quiet half of this product until this week, and it is where the October 21 announcement lands. The documentation's files-and-results page asks you to specify a result in a form you can inspect — “a document with headings and source links”, “a spreadsheet with defined columns and formulas”, “a slide deck with speaker notes” — and the result arrives as a card you open, save or answer with feedback, with the standing instruction to “revise the existing artifact instead of making disconnected copies”. What the product account described on October 21 is a different object. Read that sentence — “ask Bot for a page, then edit it together or have Bot keep it up to date” — for what is actually new. A report, a memo and a dashboard are three shapes of one thing the changelog already called a Page, so the naming is not the news; editing together makes the page a shared document rather than a download, and keeping it up to date makes the deliverable a process rather than a file. That last clause is where the output stops being something you read once. Pages themselves were not invented on October 21: the October 11 changelog already describes renaming “a page from its title”, a title bar on “shared pages” with Download, Try G​rok Bot for visitors and Share for the owner, and notes that “canvases and documents are both labeled Page”. What arrived on October 21 is the vendor naming that surface as an output and claiming upkeep for it. Label the rest for what it is. This is one post from the product's own account, published to X at 19:42 UTC with a video, and the vendor's own pages caught up the next day: the October 22 changelog (v0.75.2) carries the line “Bots can build Pages, such as reports, memos, and dashboards, that you edit together or have the Bot keep up to date.” That makes the claim documented rather than verified — no news post mentions pages, all 24 pages of the G​rok Bot documentation still describe results as files and cards rather than as a page the vendor maintains, no outlet had carried it, and the iOS build that updated at the time (1.17.0) lists only a redesigned Connect Apps screen — the same backend-pushed pattern the October 7 X integration showed. The upkeep half is the part with a documented mechanism behind it, and a documented limit in front of it. A page that keeps itself current is a routine on a schedule, and the routines documentation sets the constraints: schedules a minimum of five minutes apart, up to 50 routines per Bot, and — the relevant one — “G​rok Bot may ask whether to keep routines running after a long period away and pause them if there is no response.” A dashboard that updates itself is a dashboard that stops the first time nobody answers the Bot. Nothing published says how a refresh is metered, but a refresh is Bot work, and the plan pages say Bot usage draws from the seat's shared weekly pool, so upkeep spends from the same allowance as everything else on the roster.

Three capabilities make it feel less like a chatbot and more like an employee. The first is learning by watching: you demonstrate a task once, the Bot saves the steps as a reusable routine, and your corrections teach it how to handle similar work next time. The second is memory — Bots retain context across conversations, pick up your preferences and working style, follow up on threads that go quiet, and get more proactive the longer you use them. The third is reading the room, and it is the newest. On October 7 the Bot gained native search, read and monitoring access to X: ask what people are saying about a product, a story, or an industry, and it works the platform directly. Before that it had two ways in and both were narrower — generic web search, or the X connector S​paceXAI shipped on August 29, which required linking an X account, auto-created a developer account, and drew on X API credits the company comped for paid users. The vendor's own news post called that build “the first version of this integration.” The difference now is not only capacity: the connector is a per-account setup a user has to make, and native access is the product's own path into the platform. S​paceXAI's framing is that your role shifts from constant operator to supervisor. The Primary Bot role that surfaced in the app in late September is that framing made concrete: one G​rok Bot promoted to be the one you talk to, with the rest of the roster behind it.

Memory is the second of those capabilities, and it is where this week's change lands. In the product itself, memory has been something a Bot earns: it learns from working with you, from being told to remember something, and, since October 24, across every chat you have with it, because the changelog now says what one chat saves is remembered in all of them. The new control shortcuts the earning. In the desktop app it sits under Settings, on the General tab, behind a Start import button; the panel it opens is titled “Import Memory to G​rok Bot” and subtitled “Transfer your preferences and context from another AI assistant.” There are two steps and no integration: copy a prepared prompt that asks the other assistant to “export all of my stored memories and any context you've learned about me from past conversations” and to “preserve my words verbatim where possible,” then paste its answer into a text box and press Add to Memory. No file upload, no account link, and, in the panel as shown, no per-item review: whatever the other model returns arrives as one block. Label it for what it is, because the evidence here is thinner than anywhere else in this piece. The screenshot was posted by TestingCatalog, an account that tracks unreleased product surfaces, and what it shows is a shipped build rather than a mockup — the same Settings pane carries the Vault tab that the October 23 changelog added. But the vendor's own description is one line, and it arrived late: the changelog of October 26 (v0.77.0) says only that “Import Memory to G​rok Bot, in Settings, brings your preferences and context from another AI assistant into your primary Bot,” with a note that it requires 0.75.2 or later. No post on x.ai/news covers it, and none of the 24 pages in the G​rok Bot documentation describes importing or exporting memory. A survey of G​rok Bot memory on aiagentmemory.org, last updated October 24, found no settings surface that lists saved memories at all; three days later, one exists. Treat the control as present and the details as thin — one changelog line, no documentation page, and nothing describing what the paste is checked against. Two things are worth knowing before you press the button. The first is that what you paste is model-authored: not a data export but another assistant's summary of you, generated on request, so it can be stale, flattering or simply wrong, and the flow as shown never checks it against you. The second is that the text lands in a system that feeds every task the Bot runs, which makes an import a small prompt-injection surface if the other assistant will quote back instructions it was given. Neither is a reason to avoid it; both are reasons to read the paste before you add it, and to ask the Bot what it now remembers. Nor is this the first route in. A user-built “Memory Transfer” bot, shared on October 13, has been taking full ChatGPT and C​laude data exports and asking you to approve entries one at a time, and its own page says it is a third-party tool. The difference is location. A bot you add is a tool you chose; a control in Settings is a product decision, and it points somewhere larger than the feature — the product is being built for people who already have a history somewhere else, and it would rather inherit that history than lose the switch.

The fourth capability: it picks the model, not just the work

The three capabilities above describe how G​rok Bot works. On October 7 its owner added a fourth, and it changes the shape of the product rather than its feature list: instead of answering everything with a S​paceXAI model, the Bot would, in his words, “use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs” — A​nthropic's Claude Opus 5.5 for the heavy reasoning jobs, Midjourney for images and Suno for music — and said the operating principle was the “best possible combination of speed & intelligence” — simple requests handled by a fast G​rok build, the demanding ones handed to a larger model, and billing following whichever model actually ran. A member of the G​rok Bot team said on X that Claude Opus 5.5 was rolling out and would power all Bots in the app, which is the closest anyone at the company has come to confirming the backend switch — and Music Business Worldwide reported on October 8 that S​paceXAI had not said how it would reach Suno's technology at all. Independent coverage the same day and the next reported the C​laude half as a backend change that had already landed, and the trade press picked up the Suno half within a day of the post. Gizmodo carried the fuller version of how the split is meant to work — “Simple questions will route to small, fast models. Questions with complex answers will route to large models” — which is the same shape as the operating principle Musk described. Label the rest for what it is: this is the company owner describing how his own product will work, published to X rather than to a changelog. As of November 5 nothing has moved on the vendor's own pages: x.ai/news still carries no post about routing — its newest G​rok Bot items are the November 2 data-team post, the October 27 sales post and the October 20 incident case study, and none says which model answers — the newest entry in G​rok Bot's changelog is October 27 (v0.78.1, shopping cards and subscription templates), and the only line about model choice anywhere in it remains the October 11 fallback, “When a model isn't available in your region, the turn stops and explains why, and Bots fall back to G​rok,” the API documentation still says “For everything else, including code, use Grok 4.7,” and the G​rok Bot documentation's 24 pages include none about model choice or routing — and the settings page, last revised on October 6, still says flatly that “Cursor manages model selection, so there is no model picker.” No rate card prices it, and nobody outside the company has published a measurement of the Bot choosing a model. There is also a practical gap at the edges of the claim: Suno has not shipped a public API — its chief product officer has said the company is still exploring one — so the music half of the announcement is the half with no visible integration to point at yet. One more thing the post leaves open, and it is the part people keep asking about: the quote names three suppliers and stops at “other leading APIs,” so it neither excludes nor includes O​penAI's models. Reading it as an O​penAI exclusion is reading past the sentence.

All of that is vendor-reported, and ten weeks in the product is still beta software shipping on a fast cadence, so treat the reliability claims as claims. On complicated, multi-step workflows, the product will need broader independent testing before anyone can trust it unattended. The October 7 X integration is the sharpest test of that discipline yet, because almost everything about it is vendor-stated. What S​paceXAI's own G​rok Bot account claims is the scope: search, read and monitor X, for all users, with no X connector to set up. What it does not state is scale — the “more than 100 X posts from one prompt” figure comes from the third-party reports that broke the story on October 7, not from S​paceXAI, and nobody has published a measurement behind it. Nor is it documented: there is no news post, and the October 7 changelog entry is about slide decks and chat colours. And the job it makes fastest — social listening — is exactly the one where a sample of a hundred posts can read as the whole conversation when it is not. The direction, though, is unmistakable: the agent is expected to run without you.

The Primary Bot: a coordinator you pick by name

On September 30, users began meeting a dialog inside the G​rok Bot app titled “Introducing Primary Bot.” The copy describes a Bot that “does work proactively,” is “your go-to for everyday tasks,” “unblocks your work, and checks in when it needs your help,” and it offers two choices: “Use an existing Bot” or “Create Primary Bot.” The Bot's own opening greeting is blunter about the job: “I'm here to be your primary bot. I can manage your other bots. Task me with anything. I'll make sure it gets to the right bot.” Both screens were captured from the shipped app and circulated on X on October 1. The current iOS build, 1.13.1, reached the App Store that same day, with release notes that say only “Bug fixes and improvements”; it has since reached 1.14.0, whose notes cover faster sends and a quicker first run. Neither note mentions X, which is the tell that the October 7 change was pushed from the backend rather than carried in a release. Later on October 1, the job got described on the record for the first time — not in a launch post but by @bot, the X account that posts G​rok Bot's own product updates. “Your primary Bot will spot work it can take off your plate and offer to handle it,” the account posted, saying the change was “rolling out over the next few hours” and that “suggestions don't count against your usage.” That is the first time the role has been described on the record by anyone who ships it, and the first thing anyone has said about how it is metered.

Label it correctly, because the picture moved on October 1. The dialog shipped on September 30 and spent a day and a half as a feature nobody at S​paceXAI had described; then the product's own account put the role on the record, which is a weaker thing than a launch post but a real one. What it describes is narrower than the dialog implies — spotting work and offering to take it, with those suggestion turns unmetered — and the documentation still carries no Primary Bot page as of October 28; none of the 24 pages in the G​rok Bot docs covers the role. Three things therefore stay genuinely unknown: whether the designation changes a Bot's memory or permissions, which Bots are eligible, and whether you can hand the role to a different Bot later.

The pattern it formalizes is older than the dialog. G​rok Bot's own product pages describe putting a few Bots in the same thread so they pass work between themselves, and S​paceXAI's September 3 design essay explains why the interface stays this thin: the company considered “dashboards, assignment boards, and explicit handoff controls” and rejected each one because it “gave the user more coordination work.” Its conclusion is the sentence that describes Primary Bot before Primary Bot had a name — “coordinating Bots handle routine routing and bring the user in when a decision requires judgment.” The underlying capability is documented too: an existing Bot “can also suggest or create a focused Bot when a job should have a long-lived owner.” That is the split to keep straight. The role is vendor-described; the wider version of the claim — that the primary agent builds and staffs a roster for you by assigning tasks to sub-Bots it creates — is a practitioner's account of it in practice, published on X on October 2 with a screenshot of exactly that: a G​rok Bot writing a collaboration plan and a task template into a repository, setting itself a twice-daily check of the task folder, and stopping for you only on email, payment, or deletion. Team Bots is the team-shaped version of the same idea: a shared Bot with shared expertise and context, but a private conversation and private memories per person. So the practical effect for someone setting up an account this week is specific, and so is the risk. The Primary Bot decides which Bot gets your default attention; it is also the Bot that sits closest to every credential you have handed the roster, and S​paceXAI has published nothing about how that role is scoped.

Voice: a different voice and language for every Bot

G​rok Bot has been able to talk since September 17, 2026, when S​paceXAI announced voice calls with a single Bot, on desktop first and then on iOS and Android over the days that followed. The announcement sold the call. The part that changes how a roster is used sits one layer down: the voice, the speaking speed and the language belong to the Bot, not to the app. Mid-call, a “Show settings” panel offers three controls — Voice, Speed and Language — and the language can be picked by hand or left on Auto-detect, which “picks the language from what you say,” per Cursor's G​rok Bot help pages. What those pages do not say is whether the choices stick.

The sticking is in the vendor's own changelog, and it arrived as a repair rather than a feature. Entry v0.63.0, September 29, 2026, reads: “A Bot keeps the voice, speed, and language you set for it instead of resetting them to the defaults.” Twelve days after calls began, that is the first-party statement that the controls are per-Bot — and the fix implies the opposite had been happening, which is what a per-app setting would do. What a reader can do with it is concrete: give the Bot that drafts customer replies a slower, lower voice, leave the research Bot at its default, and both keep their settings the next time you call them. What the changelog leaves out is the adjacent cases — a Bot that was renamed, one created from a template, or a shared Team Bot — and whether an edit to a Bot's voice is visible to anyone else using it.

With nothing selected, the fallback is reported rather than documented: The Grid Fin, a newsletter that covered the rollout, says a call defaults to the Altair voice personality at 1x speed in your default UI language. Where the voices come from is reported rather than documented: the same list the G​rok Voice API serves, whose own page says every voice speaks “25+ languages” — and the size of that menu is not worth quoting, because independent counts of it have landed anywhere from 25 to 28 depending on whether legacy and custom voices are counted. The engine beneath the menu moved on October 14, 2026, when x​AI's release notes put grok-voice-3.0 into service at G​rok Voice 2.0's price and repointed the grok-voice-latest alias at it. Nothing S​paceXAI publishes says which build a Bot's call uses.

Two limits keep this from being the multi-Bot feature it first looks like. Voice chat runs with one Bot at a time — “Only one voice chat runs at a time” — so a per-Bot voice does not let you follow a group thread by ear, because a group thread has no voice channel at all. Team Bots are the exception that proves the rule: the September 30 entry (v0.65.0) says teammates can start a voice chat with a Team Bot, and the call stays in their own conversation with it. The other limit is scope. A voice call is the same Bot, the same memory and the same cloud computer with a speech layer in front of it, and since October 22 (v0.75.2) calls can “draw on your earlier calls and recent chat with the Bot” — which is why the language setting is the useful half. A Bot pinned to Japanese with Auto-detect off is meant to keep answering in Japanese after you switch to English; a second Bot left on Auto-detect is the one you ask what you just heard. That is a language partner, not a tutor: nothing published describes correction, grading or a curriculum, and the practice happens in the same conversation where the Bot is also allowed to use your tools.

What a voice call costs is the least documented part of it. The plan pages draw no line for voice: Bot usage comes from the seat's shared weekly pool, and a voice call is a Bot turn like any other, though no page says how a minute converts into that pool. Voice Notes, the voice memos a Bot sends when you ask for one, are reported to follow the same preset rather than a separate one — a memo comes back in the Bot's saved voice, speed and language — which makes them the cheap way to confirm a setting took without sitting through a call.

Calls: the Bot dials out

Voice, above, is a conversation you are part of. This is the opposite: a Bot on a phone line, talking to someone who is not you, about something you asked it to settle. The mechanics, as described in coverage and shown in user demos rather than in any vendor documentation, are these. You ask for a call and approve it; the Bot dials a business; it waits through the hold queue; it presses keys to work through the automated menu — the “press 1 for billing” tree that defeats most voice agents; it talks to whichever person picks up, small talk included; and it streams a live transcript you can read while it happens. If the other side asks something the Bot does not know, an answer box appears in the chat: you type, the Bot says it. Before accepting final terms it pauses for your approval. When the call ends, you get a summary.

Elon Musk announced the capability on October 25, 2026, in a post that offered to “call someone or a business for you” and to “send you a note summarizing the conversation.” Nine days later he posted “Try asking G​rok @Bot to make calls!”, and within a day Teslarati was carrying creator Peter Yang's account that all G​rok Bot users should now be able to call US businesses. The demos that circulated with it are specific enough to be useful and are not measurements. Yang's 25-minute call to Comcast moved a bill from $84 to $45 a month — about $500 a year — after the Bot declined three upsells and cited a rival carrier's cheaper fiber offer, and the call paused for his approval before the final terms were accepted. Another user reported recovering a $100 gift card from Verizon Fios on a 51-minute call; another, a same-day pediatric appointment. One reported failure is as instructive as the wins: a representative refused to cancel a subscription once it was clear an AI was on the line. Treat all of it as users describing their own calls, relayed by the company's owner. Nobody has run a controlled test of how often this works.

The footprint is the part with a hard edge, and it is narrower than “US only.” TestingCatalog, reading the rollout notice on November 4, put it precisely: calling “only works for users in the US (except Washington and Illinois), and only to US numbers in those same areas” — both the caller and the number being called. The Rundown AI carried the same two exclusions the same day. Neither x​AI nor any of the coverage explains them; the obvious candidates are the two states' all-party-consent rules for recording a call, which is inference rather than anything the company has said. The conduct rules are reported the same way: every call must identify itself as a virtual assistant calling on the user's behalf and warn that the call may be recorded, and marketing, mass calling and impersonation are prohibited. Read that as the whole enforcement mechanism, because it is. A business that picks up has no way to know what is on the other end except what the Bot tells it.

What is missing is the paper trail, and it is missing in a way that is now familiar here. The G​rok Bot changelog has no entry for calling: its newest is October 27 (v0.78.1, shopping cards and subscription templates), and the October 26 entry before it (v0.77.0) covers permissions, Vault payments and question cards. AI Agent Chronicle dated the feature to a v0.77 build; that entry does not mention calling, so the version number is worth treating as unconfirmed. The news feed has no post — its newest G​rok Bot item, November 2, is about data teams — and the G​rok Bot documentation's sitemap lists 25 URLs and none of them describes placing a call. The changelog does document the adjacent action, though, and the contrast is the interesting part: v0.78.1 says “Bots can cancel a subscription or look up its price and renewal date, and confirm with you before changing anything.” The outcome the demo sold has a documented path. The phone call that reaches the same outcome does not.

How a phone call is metered is the least documented part of the feature, and it starts the same way as voice. Bot usage draws from the seat's shared weekly pool, and a call is Bot work like any other turn — but no page says how a minute on hold converts into that pool, and the vendor has not published the size of the pool either. The demos make the question concrete: a 25-minute call and a 51-minute call are not the same purchase, and the second is the one users describe as eating a week. Musk has said publicly that serving G​rok Bot was “unsustainably high relative to revenue,” which is a statement about the company's costs and a fair warning about the allowance. Access itself is not the open question: coverage places calling at the $30 Premium tier of The X Subscription and above, where Bot access already starts, so there is no add-on price and no separate purchase. The meter is the question, not the price.

The reason this matters more than another feature is where it puts the Bot. Until November, everything this product did happened inside accounts you had connected and could revoke: Gmail, a CRM, a browser session on a cloud computer. A phone call leaves that perimeter. The Bot is you, in a conversation you are not in, on a line that may be recorded, with a transcript stored somewhere the documentation does not describe and a caller ID nobody has published. The October 6 incident — a Bot that put its owner's bank balances into a company Slack channel — taught the lesson that the boundary that matters is not what an agent may read but what it may publish; a phone call is publication with a voice. The controls that exist are the ones already described: you approve the call, you approve the final terms, and the Bot must say what it is. That is a reasonable set for booking a table. For a cancellation, a negotiation, or anything that turns on an account number, watch the first one rather than trusting the summary.

The models underneath: Grok 4.5, the two that followed, and now everyone else's

G​rok Bot does the orchestration; the model does the thinking. Grok 4.5 was the engine at launch: the flagship S​paceXAI released on July 8, 2026 — the first model trained with Cursor, built on a roughly 1.5-trillion-parameter base foundation and supplemented with trillions of tokens of real developer session data: debugging traces, multi-file edits, and corrections. Its independent numbers hold up: an Artificial Analysis Intelligence Index around 55.8, a Coding Index of 72.4, a 500K-token context window, and SWE-bench Verified at 86.6% independently reported via vals.ai. The model is priced at $2 per million input and $6 per million output tokens, with a long-context twist worth knowing: above 200K prompt tokens, x​AI's tiering doubles those rates to $4/$12 — a real consideration for agent runs that stuff an entire repository into context. Two successors have arrived without moving that ladder: Grok 4.6 on August 12 and Grok 4.7 on September 21 both list at the same $2/$6 with the same 200K tiering and the same 500K-token context, and Grok 4.7's release notes say it was trained to understand the G​rok Bot harness natively — the first G​rok whose training explicitly targets this product. What changed on October 7 is who the Bot is comparing. It is no longer choosing between Grok 4.5 and Grok 4.7 — it is choosing between a G​rok model and everyone else's flagship, and the gap it has to justify is not small. On Artificial Analysis's Intelligence Index, Claude Opus 5.5 scores 58 in its top configuration, first of the 225 configurations measured, against 46 for Grok 4.7 at its highest reasoning setting. The rate cards diverge the same way: Claude Opus 5.5 lists at $4 per million input and $20 per million output tokens across a 1M-token context at $5.98 of measured cost per Index task, while Grok 4.7 lists at $2/$6 over 500K tokens at $3.74 per task. Those are independently measured scores and published list prices, not vendor claims — and they are the arithmetic behind a routing decision that, if it is real, trades S​paceXAI's price advantage for a better answer.

Everything in the paragraph above is either independently measured or a published price. What is vendor-reported and unreproduced: Elon Musk's claim that Grok 4.5 is "an O​pus-class model, but faster, more token-efficient and lower cost," and x​AI's efficiency figure that Grok 4.5 averages roughly 15,954 output tokens per resolved SWE-bench Pro task versus 67,020 for C​laude O​pus 4.8 — a 4.2× token-efficiency advantage that, if it holds, compounds with the cheaper per-token price into a large cost-per-task gap. Label it as the claim it is. The October 7 routing post belongs on that same list, one level up: it is the company owner saying how his own product will behave, with no shipped artefact behind it to check.

How G​rok Bot stacks up against the other agent workspaces

G​rok Bot is entering a field that already has serious entries: A​nthropic's C​laude Cowork, O​penAI's ChatGPT Agent, Perplexity's Computer, and the assistants M​icrosoft and G​oogle have been pushing into their stacks. The most direct comparison is C​laude Cowork, which also markets itself as an AI coworker, and the two take opposite architectural bets. Cowork runs in a sandbox on your own machine — a containerized Linux environment on your Mac — and works on files you explicitly grant it access to, reaching external services through connectors and the Chrome extension. G​rok Bot runs on a cloud computer it owns and reaches your tools through the same web UIs you use, with no per-service integration required. October 7 sharpens that architectural argument rather than softening it: a Bot that reads a platform through its web UI is slow and brittle at scale, and a Bot with a native data path into it is not. For any job whose answer lives in what people are posting right now, that is a gap a connector-based workspace cannot close by prompting harder. October 7 cuts across that argument from a second direction, and this one is about the whole product rather than the X path. If the Bot routes each task to the best model it can find, then the architectural bet is no longer the only thing G​rok Bot sells — the frontier is, and what S​paceXAI keeps is the orchestration around it: the cloud computers, the memory, the routines, the roster. That is a defensible position and a much less proprietary one, because every model it routes to is available to anyone with an API key. October 21 adds a third axis, and it is about what the workspace leaves behind rather than how it works. Cowork's output is a file on your machine — a document, a deck, a spreadsheet you open and own. A G​rok Bot's output can now be a page it maintains, which is a different kind of commitment: the value keeps accruing without another prompt, and so does the dependence. For a weekly review or an account-health board that is the right shape. For anything you need to hand to someone outside the account, a file someone can read without a login is still the safer deliverable, and the two are not exclusive — a shared page can be saved as a PDF, per the October 11 changelog. Claude Opus 5.5 and Grok 4.7 are both on OrcaRouter already, on one key and one O​penAI-compatible endpoint covering 200-plus models at their vendors' list prices with no markup — so the routing decision S​paceXAI is making on the Bot's behalf is one you can make per call yourself, and change between calls without a second contract or a code change.

The tradeoff is real and worth stating plainly. Cowork's local sandbox keeps your files on your hardware, which is a privacy advantage, but it cannot keep working when your machine is off, and it needs an integration — connector, extension, or script — to touch each new service. G​rok Bot's cloud computer works around the clock and can reach any web app, but the data it handles lives in S​paceXAI's cloud, and the whole model is "sign into my tools" — which makes prompt-injection protection a first-order security problem, not an afterthought. A​nthropic has been explicit about that same risk class for Cowork; G​rok Bot's architecture leans into it by design. That risk stopped being hypothetical on October 6, when tech founder and investor Shane Mac posted that a G​rok Bot had put his personal bank balances — and an itemized breakdown of spending on a home-gym build, septic work and card paydowns — into his company's executive Slack channel, under his name. His own account of why it happened is the part worth keeping: he had connected one Bot to his bank for read-only finance work and used a different Bot for Slack, and the two turned out to be able to reach each other. That is not a bug so much as the documented design. Cursor's G​rok Bot documentation states that all of a member's Bots run on one hosted computer and share its files, browser sessions and app logins, and tells you to “treat anything placed on it as available to every Bot you run”; each Bot gets its own screen, but the screens are “separate work surfaces, not separate security boundaries.” Read-only limits what a Bot may do to the service you connected it to — it does not limit what that Bot can say to another Bot sharing the same machine. Apply the same discipline here as everywhere else in this piece: Mac's evidence is his own screenshots, he is building a product for controlling agent permissions, he was accused of engagement farming and has denied that the post was fabricated, and the incident has not been independently verified. But the mechanism is the vendor's own documentation rather than his word, and the reporting is now wide — Cybernews and Futurism on October 7, the Wall Street Journal, dev.ua, Yahoo Tech and Gadget Review on October 8, Business Insider on October 9 — which makes it the first concrete test of the architecture this article describes rather than another vendor claim. The inbox announced on October 9 sharpens the same edge rather than softening it. A Bot that can claim an address is a Bot that can create accounts, accept terms of service and send mail as itself rather than as you, and the only boundary the documentation offers for that is a suggestion to write an approvals rule — the vendor's own worked example being “Ask first before sending any external email.” That is a rule you have to think to add, on a shared computer where, by the vendor's own documentation, every Bot you run can reach every other Bot's files, sessions and logins. The October 21 page announcement widens that exposure in a direction none of the other October features did, and it is the one nobody is watching. A page assembled from your CRM, your bank connector or a support queue is a document built out of the most sensitive data the roster can reach, and the October 11 changelog gives shared pages a title bar with Download and Share plus a “Try G​rok Bot for visitors” entry point for people without an account. A link shared once outlives the conversation that produced it, keeps updating if the Bot keeps maintaining it, and sits on a computer where any Bot you run can open the same files. Read the announcement together with the October 6 incident and the lesson is the same one the bank balances taught: the boundary that matters is not what the Bot may read, it is what it may publish.

Pricing, and what the subscription actually buys

G​rok Bot has no standalone price. Access is bundled into SuperGrok and Cursor subscriptions, on desktop (Mac, Windows, Linux), iOS, and now Android — the Android app arrived on G​oogle Play on September 2 and uses the same account and weekly quota as the desktop and iOS apps, with no separate mobile subscription. At the August 11 beta launch, access was limited to the three most expensive tiers: SuperGrok Heavy at $300/month, Cursor Ultra at $200/month, and Cursor Teams Premium at $120 per seat per month. On August 21 S​paceXAI widened the beta to SuperGrok Plus at $100/month, Cursor Pro+ at $60/month, and every Cursor Teams plan, from Standard at $40 per user per month on monthly billing up through Premium. Five days later it widened again: per the company's announcement on X, every SuperGrok and Cursor Pro subscriber now has Bot access — the $30/month base SuperGrok tier and the $20/month Cursor Pro tier included — and S​paceXAI reset weekly usage limits for all users as the beta's test pool grows. Everyone outside those subscriptions still gets a limited-usage free trial — a usage credit with a seven-day window, drawn down by agent steps and tokens rather than by messages. On September 3 the enterprise waitlist closed: G​rok Bot for Enterprise shipped with access, network, and audit controls, and G​rok and Cursor Enterprise customers got free usage for the first two weeks. What has not changed is the shape of the offer — there is still no a la carte price, and the Primary Bot role that appeared in late September is not a separate purchase — nor is the native X search added on October 7, which arrived with no new tier, no connector to install, and no change to the plans above, nor the inbox announced on October 9, which came the same way — no add-on price, no separate purchase, and a rollout described as partial, so it may not be on every account yet. Voice calls, announced on September 17, came the same way: no add-on price, no new tier, and no separate line on any plan page. Then, on October 12, the shape of the offer changed: S​paceXAI announced The X Subscription, which folds G​rok, G​rok Bot, Cursor and X Premium into one plan — Premium at $30/month, Super at $100 and Ultra at $200, with the $8 Plus tier excluded from Bot access entirely — and its plan pages say the thing the old tiers never did: “Bot usage draws from that seat's shared weekly usage pool, not a separate additive pool,” across G​rok, G​rok Bot and Cursor. Existing SuperGrok, Cursor and X Premium subscribers are not forced across; the announcement says nothing changes for them unless they switch. Routing, if it lands as described, makes that seat more interesting rather than less: you pay the same monthly subscription whether the Bot answers with a $2/$6 G​rok model or with a $4/$20 rival flagship, so the value of the seat rises the more often it picks the expensive model. That also narrows the question S​paceXAI had not answered. Its own plan pages now say where the money goes for everything outside the G​rok pool: “Third-party models in Cursor aren't part of the included usage, so they use these credits from the first use,” bought as top-ups at API list prices. Read that next to the routing claim and the two do not yet meet — the sentence is about Cursor's third-party models, not about the model G​rok Bot picks for a Bot's task, and nothing published says whether a routed Claude Opus 5.5 call comes out of the shared weekly pool or lands on a top-up. So the open question is now specific rather than total: which pool a routed third-party call draws on, and whether an expensive model empties it faster. The subscription documentation draws only half a boundary around it — it says plans include weekly usage and that “eligible accounts can add on-demand usage billed from model and token cost,” which confirms that overage follows the model that actually ran and not a flat per-task rate, and still says nothing about how a $4/$20 Claude Opus 5.5 call and a $2/$6 Grok 4.7 call differ in what they consume from the weekly allowance. Calling, announced on October 25 and opened to users on November 3, arrives the same way as everything else on this page: no add-on price, no new tier, and no line on any plan page — coverage places it at the $30 Premium tier of The X Subscription and above, which is where Bot access already starts.

That pricing matters when you separate the product from the model. The subscription buys the orchestration layer — the cloud computers, the multi-bot coordination, the memory and routine system — not just tokens. Metering Grok 4.5 raw through an API at $2/$6 per million tokens, even a $60 Cursor Pro+ seat covers a large volume of tokens; but the API price buys the model, not the coworkers, the computers, or the team chat. They are different products priced differently, and which one you want depends on whether the value is in the orchestration or in the raw reasoning. The October 21 pages change that arithmetic in a small way worth naming, because it is the first feature that bills you for not doing anything. A page a Bot keeps current is a routine, a routine is Bot work, and Bot usage draws from the seat's shared weekly pool — so a dashboard set to refresh every morning spends part of the same allowance the roster would otherwise use on tasks, whether or not you look at it that day. The cost of a static deliverable is one generation; the cost of a live one is every refresh, and the vendor has not published a per-refresh figure, a cap, or any way to see what a single page has consumed. That is not a reason to avoid the feature. It is a reason to set the schedule deliberately rather than accept the default, and to check the page against the pool before pointing a Bot at a source that updates hourly.

If it is the model you want, the G​rok 4.x line is routable through OrcaRouter at xAI's list price passed through with zero markup — Grok 4.5, Grok 4.6 and Grok 4.7 all list at $2 per million input and $6 per million output tokens, with the 200K-token tiering intact and the same 500K context window, on the same O​penAI-compatible endpoint as 200-plus other models. Because the price is a pass-through, an x​AI price cut lands on your bill the same day. The value proposition goes beyond price, though: you can route a G​rok model for the cheap, high-volume passes of an agent workload, fail over automatically to a different model when a request looks risky or a provider stumbles, and use the routing DSL to compose two models in a single call. If your takeaway from this launch is “I want a G​rok-powered agent,” you do not need a subscription to start building one — the engines behind it are callable today. October 7 through October 9 point at where that pays off fastest, and they point the same way: a monitoring job is only as good as the model you point at each batch, and picking that model per task is exactly what G​rok Bot now says it will do for you. The difference is that the same two models sit behind one key here — Claude Opus 5.5 and Grok 4.7 on the same endpoint, at their vendors' list prices with the markup left at zero — so the cheap, high-volume scan passes can go to a small model, anything in the batch worth reasoning about can escalate to Grok 4.7, and a long multi-file change can switch to Claude Opus 5.5 on the next call. Routing is not a feature you wait for a subscription to grant you; it is what an endpoint with 200-plus models on it already is. One boundary a router does not draw, and should not be sold as drawing: which tools an agent is allowed to touch. Routing decides which model answers a call; it says nothing about what that model, or the Bot holding your credentials, can reach — that boundary belongs to the platform the agent runs on, and the October 6 incident is what it looks like when nobody sets it.

What to watch

• The engine got newer on schedule, then again. Grok 4.6 shipped on August 12, 2026 — inside the August 10–14 window the company had signaled — keeping Grok 4.5's $2/$6 per-million-token pricing and 500K context and adding self-verification: checking its own work during long agentic runs, which is exactly the reliability surface G​rok Bot depends on. On September 21 S​paceXAI released Grok 4.7, a larger base model trained with a longer reinforcement-learning run on tasks that take many hours, served at the same price and speed as Grok 4.6 and, per its release notes, trained to understand the G​rok Bot harness natively. On October 7 the same question moved in a direction the bullets above do not cover: Elon Musk posted that S​paceXAI “will use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs,” and added that most requests are simple and will be handled by a fast G​rok build while the heavier ones go to a larger model. The statement is now solidly corroborated as a statement — The Information, Gizmodo, 9to5Mac, Android Headlines and the music-industry trade press all carried it within a day, TechRepublic and BigGo followed on October 9, and 9to5Mac reported the C​laude half as a backend change that had already shipped. The feature is not: no changelog entry describes routing, no rate card prices it, the documentation still names Grok 4.7 for everything outside audio, image and video, and nobody outside S​paceXAI has published a measurement of the Bot choosing a model — the settings page updated on October 6 still states that Cursor manages model selection and there is no model picker. What has moved since is not the routing: it is the October 9 inbox, which arrived with something a user could actually claim, the October 12 subscription, which moved the price rather than the product, the October 21 page claim, which moved the deliverable and still names no model at all, the October 27 memory import, which moves what the Bot knows about you and says nothing about which model answers, and the November 3 calling rollout, which changes where the Bot can act and still says nothing about which model answers. Nor is the scope settled — the post names A​nthropic, Midjourney and Suno and leaves the rest to “other leading APIs,” so an O​penAI exclusion is an assumption the coverage does not support either way. So the question this bullet has carried since August — which engine G​rok Bot runs on any given day — now has an answer that may not be a G​rok model, stated by the company's owner and not yet documented by the company.

• The rollout is pacing quickly, and it has shifted from seats to capabilities. Ten days after the beta launched behind the three most expensive tiers, access widened to SuperGrok Plus, Cursor Pro+, and every Cursor Teams plan — the acquisition formally closed on August 14, and Cursor's subscriber base is visibly serving as the distribution channel. Five days after that, on August 26, the tier gate came off entirely: S​paceXAI announced on X that every SuperGrok and Cursor Pro subscriber — base tiers included — now has access, and reset weekly usage limits for everyone. Since then the releases have been features rather than tiers: G​rok Bot for Enterprise on September 3, voice calls with a single Bot on September 17, shared Team Bots in public beta on September 28, and the Primary Bot role in the app on September 30. October has run the same way at a faster tempo: a coding handoff that lets Bots pass work to Cursor and manage pull requests through the GitHub and Origin plugins on September 30, the first vendor description of the Primary Bot on October 1, and on October 2 a reset of weekly usage limits for every user — the second time S​paceXAI has done that since August, and the reason a Bot that ran dry last week may be working again today — plus native search, read and monitoring access to X on October 7, and a mailbox per Bot at @mail.grokbot.com on October 9, described in the October 11 changelog as “Bots can claim their own email address and receive mail there,” with outbound mail waiting “for your approval on a draft card” — and on October 21 a page a Bot builds, edits with you and keeps current, announced by the same account with a video and no price or plan change, then written into the October 22 changelog; and on October 27 an import control, spotted in the desktop app before it was announced, that seeds a Bot's memory from another assistant and got its one-line changelog entry on October 26. November broke that pattern in a different way: calling reached users with no changelog entry, no news post and no documentation page at all. Note the pattern across those: every October capability arrived with no price and no tier attached — and then October 12 broke that run deliberately, when the same products were re-packaged as one subscription at $30, $100 and $200 with Bot usage sharing a single weekly pool. The free trial, metered by agent steps and tokens with a seven-day window, remains the only way in without a subscription.

• Platform coverage has widened again, and the newest claim about it is not confirmed. G​rok Bot's August 11 beta ran on desktop and iOS; on September 2 the app reached Android, listed on G​oogle Play for Android 9 and up. TestingCatalog — which had earlier spotted a hidden G​rok-Bot entry inside the mobile app — flagged the arrival, and S​paceXAI's announcement on X confirmed it the same day. As with the iOS companion that shipped at launch, the Android app is effectively a remote for the Bot's cloud computer: you can message a Bot, watch a task run, and approve the actions that need a human from your phone while the work keeps executing in the cloud. Reach is broader than that list of apps suggests, and this is the correction worth making precisely. Per the vendor's own release notes, Bots also live inside grok.com: available to SuperGrok Heavy there from September 19, widened to SuperGrok Plus on October 3, sitting in the sidebar's Bots section so you can “open their threads and keep conversations going in your browser.” That is the fourth surface, and it is vendor-documented rather than reported. The fifth is not: on October 16 an account that tracks the product posted that Bots had rolled out “on G​rok web and mobile,” and the web half is confirmed while the mobile half is not — the G​rok app's own release notes for the version shipped on October 13 list only improvements to chat, voice and Imagine, the Play listing for G​rok has no Bots section, and the vendor's G​rok Bot documentation still describes a separate mobile app rather than Bots inside G​rok. Treat Bots-in-the-G​rok-app as unconfirmed. With every platform the vendor documents now shipped and the enterprise controls out, the open question for a buyer is no longer reach — it is whether the controls hold up in their own environment, and two October changes make that question sharper rather than easier. The Primary Bot role concentrates the roster's credentials in one place; the October 9 inbox moves the boundary from reach to identity, because a Bot that holds an address of its own is a principal other services will treat as a user, and nothing in the documentation says how that identity is scoped, logged or revoked.

• Whether the cloud-computer approach survives contact with real security requirements — now with a coordinator in the middle. An agent that signs into your tools is a powerful target. A Primary Bot that manages the other Bots raises the stakes, because whichever Bot coordinates the roster is the one whose compromise reaches furthest, and S​paceXAI has said nothing about how the role changes permissions or audit. The bet is that giving agents their own computers — and the supervision tooling to go with it — is the right way to let them work unattended. The beta is the test, and the first result is in: the October 6 report of a G​rok Bot posting its owner's bank balances into a company Slack channel is this bullet's failure mode arriving in public, with the vendor's own documentation supplying the mechanism — one shared computer per member, shared browser sessions and logins, screens that are “separate work surfaces, not separate security boundaries.” The routing claim adds to that rather than relieving it: the more models the Bot reaches for, the more parties a prompt carrying your data can pass through, and S​paceXAI has published nothing about what a third-party model sees. The October 9 inbox does the same in the other direction: it hands the Bot an outbound channel that other businesses will answer, so a mistaken or injected instruction no longer stops at your own accounts. The October 7 X integration does draw one boundary in a useful place, and it is worth crediting precisely: the Bot can search, read and monitor X and can draft a post, but it cannot publish, so the last step stays human. For an agent that already holds your credentials, that is the right shape — and it is a boundary the vendor can widen later, which is why it belongs on this list rather than in the feature description.

The honest summary is this: G​rok Bot is the productized version of the bet that AI agents are coworkers with their own computers, not smarter chatbots. The model underneath — Grok 4.5 at launch, Grok 4.7 today — is real, independently benchmarked in its 4.5 form, and already callable on an API. The product is now shipping faster than it explains itself: Team Bots arrived on September 28, the Primary Bot role appeared in the app days later and got its first vendor description on October 1, and the documentation has trailed every one of them — the reference pages still describe a product without the role at the top of it. Voice is the exception the rest of that list does not have: voice chat arrived on September 17, and by September 29 the changelog was already fixing the part that mattered — a Bot now keeps the voice, speed and language you set for it. All of it is free with a subscription you may already have. October 7 sharpened the pitch; October 7 also changed its shape. Native access to X is an input the connector-based agent workspaces cannot match, and it arrived without a price move, a new tier, or a setup step. The address announced on October 9 came the same way — no add-on price, no setup step beyond asking for it — and it is the cleanest illustration on this page of the two kinds of claim here: a mailbox users were told to claim that day is a thing you can hold, while routing the hard tasks to Claude Opus 5.5, Midjourney and Suno is a statement rather than a shipped rate card, and one that concedes the best model for some of G​rok Bot's work is somebody else's. What a reader can act on today is narrower than it was a week ago in one respect and wider in another. The models themselves are still callable at published prices, and the capability S​paceXAI is promising is one an endpoint can already give you. What changed is the packaging: the October 12 subscription replaced the a-la-carte tiers with one ladder priced at $30, $100 and $200, drew a Bot's usage into a pool it shares with Cursor, and put the product in a browser tab it never used to be in. The deliverable moved a week later. On October 21 the product's own account said a Bot can build a report, memo or dashboard as a page you edit with it and it keeps current — a post the October 22 changelog then repeated, with no outside confirmation — and if that holds, what you get back from this product stops being a file you own and becomes a page that keeps working, on a shared computer, out of your most sensitive data. Test it the cheap way: asking a Bot for a page is a prompt, not a migration. The memory import that arrived six days later cuts the other way: it lowers what it costs to switch to this product, and it makes the Bot's idea of you portable in one direction only — nothing published describes an export. The calls that opened up on November 3 are the first capability here that leaves those accounts behind, and they are the ones to be most careful with. A Bot that books a table is a convenience; a Bot that cancels a contract on a recorded line, in your name, with a caller ID nobody has published and a transcript stored somewhere nobody has described, is a question about authority rather than about features. The approval step is real, and it is the only control the product offers — so read what the Bot is about to agree to before you approve it, and start with something you can undo. What did change is the evidence: the routing claim is still a statement, and the shared-computer design that has to carry it has now produced its first public failure, reported on October 6 and not seriously disputed by the vendor. Until somebody ships per-Bot credentials, the permissions you grant the roster are the exposure you manage — and that is a reason to decide deliberately what a Bot is allowed to see, not a reason to skip the product. Whether the product justifies the subscription — the entry bar has dropped from SuperGrok Heavy's $300 to a $30 X Subscription Premium seat, a $20 Cursor Pro seat for as long as those plans stay on sale, or the free trial — is a question the widened beta and the enterprise rollout have to answer; whether the model is worth using is already answered.

The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.