2026-08-28 · Guide
Route Cheap Work Off Grok Bot Without Inventing a Model Picker
A pile of two hundred vendor PDFs is the job that makes people invent a Grok Bot model picker, and the product still does not have one.
Grok Bot has no model picker, for members or admins, and the teams docs do not plan to allow admin or user choice (teams and enterprises). grok bot deepseek routing, as this page uses the phrase, is not "run DeepSeek inside Grok Bot." Keep Grok Bot for standing computer jobs. Send bulk summarization to a tool you already pay for, after you confirm that tool on its own site.
This page is that split. The two-phase meter lives on Grok Bot weekly allowance: an included weekly pool, then on-demand work billed from model and token cost (Grok Bot FAQ). What writes overflow lives on Grok Bot on-demand usage. Failover billing is a serving fact, not a DeepSeek menu. The Bot serving set is unpublished. grok-4.6 belongs to Grok Build.
Two hundred PDFs is an arbitrary example, not a published pool size. This article will not invent an allowance dollar.
Treat the missing model picker as a routing rule, never as a hidden Settings page
The documented sentence is the whole control. There is no model picker. There is a fixed model set per surface, with automatic failover. Billing follows the actual serving model. You do not choose it, restrict it as an admin, or limp through a PDF dump by selecting a cheaper name.
If a settings page mentions Default Model, when model selection is available, treat that as a different surface. The Bot docs still refuse a picker. How bots on Grok actually work walks that refusal. Stay here when the refusal is what forces work off the bot.
Searching Settings is how a dump still runs. You spend twenty minutes looking for DeepSeek while the briefing bot is already on PDF one. Pause is the control you have. A picker is not.
| What you opened Settings to find | What the Bot docs describe | What you do instead |
|---|---|---|
| A cheaper model for a PDF dump | No picker for members or admins, and none planned | Route the dump off the bot. Keep the standing job |
| DeepSeek as a Bot serving option | Unpublished Bot serving set. Failover you cannot aim | Confirm DeepSeek on its own site. Do not invent a Bot tile |
| A pin to grok-4.6 so the dump feels official | grok-4.6 powers Grok Build. Do not map it onto Grok Bot | Leave Build facts on Grok Bot vs Grok Build |
| An admin lockdown on model choice | Admin choice is not planned | Pause, coarsen clocks, keep cheap token work off the computer |
All bots on an account share one persistent cloud computer assigned to the user, not to a bot (computer and apps). The weekly pool follows that grain. A named bot that "only summarizes" still drinks the same tank as Inbox Triage. A second screen is not a second meter.
Keep standing computer jobs on Grok Bot and send bulk summarization off the weekly pool
Grok Bot is a named teammate on a managed Linux VM you can leave running with the laptop shut. A weekday brief, an inbox pass, a mail tidy, a clock that opens a browser and files a note under /workspace: those jobs need the computer.
Bulk summarization of a pile you already hold is token work. It needs a model call you can meter on a tool you already pay for, then a folder the Bot can read later. Leave the pile on the Bot and the Bot becomes the model call. The weekly pool takes the hit.
Chief of Staff Briefing is a standing job when it reads calendar, owed replies, and overnight changes. It becomes a pool-burner when you hand it two hundred PDFs before 07:00. Podcast Summarizer is one episode from a transcript, with a timestamp on every line, not a folder of two hundred files.
Route the pile. Keep the clock.
| Work | Keep on Grok Bot | Route off | Why |
|---|---|---|---|
| Weekday brief from calendar and overnight changes | Yes | No | Needs the computer, the logins, and a clock |
| Inbox labels and three unsent drafts | Yes | No | Needs mail on the shared computer. Still never sends |
| Two hundred vendor PDFs you already downloaded | No | Yes | Token work. A Bot pass will drink the week |
| One podcast transcript with timestamps | Maybe, as one episode | Yes if you have a batch of dozens | One episode can be a Bot job. A batch is a pile |
| Recapturing a live checkout in the browser | Yes, with a capture cap | No | Needs the computer. Still not a model-picker problem |
Cheapest paid path is Cursor Pro at $20 a month, checked 23 September 2026 against cursor.com/pricing. Every paid Cursor plan includes Grok Bot, from Cursor Pro at $20; Cursor Hobby, the free plan, does not, and an individual SuperGrok, SuperGrok Plus, SuperGrok Heavy or X Premium+ subscription can be linked instead. Seat price is the door, not a picker, and not a published allowance. Confirm SuperGrok Plus, Heavy, Ultra, and Teams on x.ai/pricing and Cursor's live pages.
Hedge DeepSeek on its own site before you treat it as the cheap lane
DeepSeek is a name operators type when they want grok bot deepseek routing to mean a cheaper model inside the Bot. That meaning is false. This page will not print DeepSeek prices, context windows, or feature lists as facts. Confirm the current product and invoice on DeepSeek's own site, and on the site of whatever summarizer you already pay for.
If you do not already pay for a summarizer, do not open a new vendor because a roundup said it was cheap. Cheap against an unpublished Grok Bot pool is a comparison you cannot finish. Compare against an invoice you already have, confirm it, then file the output.
A Bot that curls an API from the shared computer is still a Grok Bot turn. The unpublished serving model still ran. The weekly pool still moved. Routing means the human (or a job outside Grok Bot) does the pile, then a person drops files into /workspace.
Hosted MCP sign-in tokens stay with Cursor's backend, not on the computer (teams and enterprises). That sentence is not a DeepSeek connector. Do not invent an MCP tile. Do not paste an API key into a Bot charter and call the paste a picker.
Count two hundred PDFs as an arbitrary pool-burner, never as a Bot charter
Two hundred is an arbitrary count so the failure is obvious. Forty decks or nine hundred scans follow the same rule. If the work is "read this stack and write shorts," and the stack already sits on disk, it leaves the Bot. If the work is "open this site every weekday and file what changed," it stays.
There is no Grok Bot-specific spend cap, only the account-level On-demand monthly limit. Overflow is on-demand from model and token cost if the plan allows it. A PDF dump does not meet a dollar slider. It meets an empty week, then a bill, or a stop you did not plan. Grok Bot cost is how frequency multiplies. This page is the pile you should not put on that roster.
A routine assigns a workflow to one bot. Max 50 routines per bot. The app keeps the 20 most recent run records per routine. Deleting a bot deletes its routines. Nothing is team-level (skills, routines, and automations). Twenty records will not reconstruct two hundred PDF passes. Individual accounts and self-serve Teams still have no audit view of Bot actions; Enterprise has audit logs and Action Recording.
Do not schedule the dump. Teach-by-demonstration records visible computer interaction for up to ten minutes, browser workflows only, no microphone, draft skill, unavailable on iPhone. A PDF pile is not that recording. Opening Acrobat on camera does not create a cheap model.
Watch Ivo empty the week on a PDF dump a cheap API would have finished
Ivo is an invented operator, not a real person and not a customer. On Monday 25 August 2026 he copied two hundred vendor PDFs onto the Grok Bot computer and told Chief of Staff Briefing to summarize every file overnight. He searched Settings for DeepSeek when the first files looked slow, found no model picker, and left the job running.
Tuesday morning the brief was thin, the weekly pool was empty, and the dump was still walking files. Overflow had started, billed from model and token cost, on whatever serving model actually ran. Ivo never picked that model.
| Clock | What Ivo did | Pool in this story | What should have happened |
|---|---|---|---|
| Monday 18:10 | Copied two hundred PDFs onto the computer | Inside the weekly pool | Leave PDFs on a laptop. Do not hand them to the briefing bot |
| Monday 18:25 | Searched Settings for DeepSeek, left the job running | Still inside | Pause. Route the pile off the bot |
| Monday night | Briefing bot read PDFs until the pool died | Pool empties | Human batch on a cheap API, confirmed on that vendor's site |
| Tuesday 07:00 | Brief missing. Overflow on. Inbox clock still armed | On-demand | Pause every routine. Stop overspending |
| Tuesday 21:40 | Off-bot pile, shorts dropped under /workspace/vendor-shorts/ | Overflow already billed | The drop is correct. It does not refund Monday night |
| Wednesday 07:00 | Briefing bot read the shorts folder, not the PDFs | Standing clock | grok bot deepseek routing as a human split, not a Bot menu |
Ivo's inbox bot was still on a weekday clock. All bots share the computer and the pool, so the dump burned the account, not only the briefing bot. Pause would have stopped both. A picker would have stopped neither.
On iPhone he could have paused. He could not have invented a model menu. Editing and testing a routine still need the desktop app; the phone can now show run history and delete a routine (mobile). He did not pause. The night still billed.
Drop finished summaries into /workspace instead of asking the bot to re-read the stack
The computer is the place results land, not the place the pile must be read. After the cheap pass finishes, copy the shorts into a folder the standing bot already uses. Name the folder, the file shape, and the date on each file. Then the briefing bot reads fifty shorts, not two hundred PDFs.
/workspace is shared. Cookies, sessions, files, and command-line credentials are shared across bots. Deleting a bot does not remove those files. Screens are not security boundaries. Do not use separate Bots as a security boundary (approvals, security, and privacy). A vendor PDF pile on the computer is readable by every named bot. If those PDFs should not sit next to a mail session, keep the pile off the computer and drop only the shorts.
The drop can be markdown, CSV, or a sheet export. Pick one shape. The standing bot should refuse to open *.pdf in the vendor folder once shorts exist.
Do not ask the bot to match shorts against PDFs by re-reading the PDFs. That is the dump again. Spot-check ten files yourself. If ten fail, fix the off-bot job.
Refuse to invent a Bot model menu, a SKILL.md pin, or a named serving model
Three forgeries show up in the same hour as the Settings search.
Forgery one is a Bot model menu. A prompt that says "use DeepSeek for PDFs and the default model for mail" does not create a picker. The serving set is still fixed. Failover is still automatic. Billing still follows the serving model you did not choose.
Forgery two is a SKILL.md pin. Claude Code, SKILL.md, and CLAUDE.md compatibility is Grok Build, never Grok Bot. The Grok Bot docs never mention those files. Grok Build accepts but does not apply SKILL.md model anyway. Pasting a skill into a Bot charter because a tweet said Grok is compatible with Claude Code is a Build fact applied to the wrong product. Details live on CLAUDE.md, skills, and MCP.
Forgery three is a named serving model. grok-4.6 is real, knowledge cutoff 1 February 2026, and powers Grok Build in the developer docs. Do not map it onto Grok Bot. Do not map DeepSeek onto Grok Bot. Do not write a hyphenated stack name to make the dump feel configured.
Privacy Mode (Legacy) blocks Grok Bot entirely. That is not a cheaper model. That is the product off.
Split cheap token work from browser work before you attach a weekday routine
Attach a routine only after the pile is off the bot. Max 50 per bot is a cap on cards, not a spend cap. A daily brief that reads /workspace/vendor-shorts/ is a computer job. A daily brief that globs /workspace/vendor-pdfs/*.pdf is the dump on a clock.
Browser work stays. Token piles leave. Mail Cleanup Assistant still needs the inbox on the computer. YouTube Transcript Desk should use a transcript plugin rather than burn the browser on captions, and it still files one brief. Two hundred videos is another pile: batch off the bot, drop shorts, let a standing bot read the folder.
Static egress IPs mean some services flag datacenter addresses. That is a login problem for browser jobs, not a reason to summarize PDFs on the Bot to "stay inside one IP." The cheap pass can run on your laptop.
Beta launched 11 August 2026. Eligibility widened 21 August 2026. The docs say choice is not planned. Do not wait for Ultra at $200 a month to grow a DeepSeek tile. Ultra includes Grok Bot. It does not include a picker.
You are Chief of Staff Briefing. You write a weekday morning brief.
You never send mail, never reply, never move a calendar event.
Read, in this order:
1. Calendar for today.
2. Replies I owe, and who owes me, from the mail the inbox bot already filed.
3. Files under /workspace/vendor-shorts/ dated since the previous brief.
Each file is a short another tool already wrote. Quote the short.
Do not open /workspace/vendor-pdfs/. Do not glob *.pdf.
Do not call an external model API. Do not name DeepSeek, grok-4.6,
or any serving model. There is no model picker.
If vendor-shorts is empty, say "no vendor shorts since last brief" in
one line and stop that section. Do not go looking for PDFs to fill the
gap. Do not retry. File the brief under /workspace/briefs/YYYY-MM-DD.md.
Write a boundary that the bot never opens the raw PDF pile
The boundary belongs in the charter. The bot never opens the raw PDF pile. It never calls an external model API as a substitute picker. It never sends. Those three lines are what make it safe to leave the standing job running after you have routed the dump.
An approval controls the proposed action. It does not reverse work already completed. If the bot already started PDF one, denying a later step does not unread it and does not refund the tokens. Pause first. Edit the charter at a desk. Drop shorts. Resume the standing clock, coarser if the week already overflowed.
Do not use a second named bot as the PDF reader "so the briefing bot stays clean." Screens are not walls. The second bot still drinks the account pool and still sees the same files. Isolation is a second eligible account, not a second name. One computer, many screens is that architecture.
Shipped since: a team-level ceiling on local execution, and, on Enterprise, an admin terminate that ends a member's computer while the durable disk is kept. Neither is a model picker. Do not wait on them to process two hundred files.
Answer the operator who wants DeepSeek to run inside Grok Bot
The strongest case for running DeepSeek inside Grok Bot is simple. The Bot already has the files and a clock. A cheap model on the same computer would keep one workflow. Routing off the bot means a human step, a second vendor, and a drop into /workspace. If DeepSeek is cheaper per token, the Bot should call it. If Settings has no picker, paste a key. If the serving set failovers, maybe it already failovers to something cheap.
Grant the files and the clock. The computer is the right place for results. A second vendor is real overhead. A human drop is a step you can forget on Thursday.
The argument still loses as a standing rule. There is no picker to point at DeepSeek. Pasting a key does not create one. A Bot turn that orchestrates an API is still Bot usage on an unpublished serving model, billed from model and token cost after the weekly pool, with no Bot-specific spend cap. Failover is not a cheap-lane selector you aim.
Where the argument wins is not inside Grok Bot. It wins when you already pay for a summarizer, you confirm it on that vendor's current page, you run the pile there, and you drop shorts onto the computer. That is grok bot deepseek routing as an operator split. If policy forbids the second vendor, the pile waits for a human, or you pay overflow with both eyes open. A forged menu is not an option.
Verify the split with a Wednesday check that can fail
A split you cannot fail is a story. Run this check the Wednesday after a dump week. If it fails, the pile is still on the bot.
| Check | Pass | Fail |
|---|---|---|
| Charter glob | Briefing bot has no *.pdf glob and no external model API | Charter still says summarize every PDF in vendor-pdfs |
| Folder | /workspace/vendor-shorts/ has dated files from an off-bot job | Folder missing, or only raw PDFs exist |
| Settings | You did not add a model name, a DeepSeek tile, or a grok-4.6 pin | Charter names a serving model |
| Clocks | Standing jobs paused during the dump, then resumed without a PDF glob | Inbox or brief still armed while a dump runs |
| Invoice | You can point at overflow from the dump night without quoting an allowance dollar | You invented a pool size to justify the dump |
| Spot-check | Ten shorts match ten source PDFs you opened yourself | You asked the bot to re-read the PDFs to "verify" |
The 20 most recent run records are not this table. Write the Wednesday row yourself. If the brief is green and the PDFs still live in the glob, the check failed. Quota exceeded is the banner after the pool is gone. Do not wait for it.
Send failover-billing readers to on-demand usage, not through this routing page
Automatic failover is real. Billing follows the actual serving model. That pair is why people search for a picker after a spike: they want to pin a cheaper name so failover cannot "upgrade" them. This page will not turn that search into a menu.
If the question is what writes overflow after the weekly pool, go to Grok Bot on-demand usage. If the question is the two-phase meter, go to Grok Bot weekly allowance. If the question is a spend cap that does not exist, go to spend cap and token burn. Stay here when the question is how to keep two hundred PDFs from becoming Bot work in the first place.
A twin page about billing following failover is not in this corpus yet. Until it is, those two billing articles are the twin. Do not file a PDF dump under failover. Failover is not a route to DeepSeek.
With both a Cursor and a SuperGrok subscription, Grok Bot uses whichever has more usage. That is a pool rule, not a picker, and not a DeepSeek flag.
Name what still belongs on Grok Bot after the summaries land
After the drop, Grok Bot still earns the seat. The weekday brief still opens mail and calendar. The inbox bot still drafts and still must not send. A competitor page that changed overnight still needs a browser. Routing the pile does not mean pausing the product.
What leaves and stays gone is the bulk read. Do not sneak it back as "just this week, the vendors sent a lot." A lot is the pile. If vendors send two hundred PDFs every Monday, the off-bot job is weekly and the Bot reads shorts daily. A Bot that eats the Monday folder is a relapse.
Six bots per channel is a hard ceiling from the practitioner guides: a Projects Manager plus five others. Do not spend a sixth seat on a PDF reader. Reuse the briefing bot after the folder exists. Create new specialists only after a human approves.
The computer is a managed Linux VM. The Bot runs as a non-root user. That is not a Linux desktop client. Supported clients are macOS (Apple silicon and Intel), Windows (x64 and Arm64), and iPhone on iOS 18 or later. There are Linux desktop and Android apps as of September 2026, and the iOS app also runs on iPad (iPadOS 18 or later). None of those facts grow a model picker.
Stop this routing page when the job is a standing computer workflow
This page stops applying the moment the work needs the Bot computer to click, wait, and file, and does not need a bulk model pass over a pile you already hold.
If a five-minute inbox clock is eating the week, go to how to stop overspending. If the mix-up is Grok Bot versus Grok Build versus a model name, go to Grok Bot vs Grok Build. If you never had an eligible invoice, you are not routing.
Keep using this page while a dump is being framed as a missing DeepSeek tile. The boundary stays: the bot never opens the raw PDF pile, never calls an external model API as a substitute picker, and never sends. Ivo's Monday night is the dated failure.
Keep reading: Grok Bot On-Demand Usage: What Burns the Extra Tokens, Grok Bot Weekly Allowance: How Overflow Billing Actually Works, Grok Bot vs Grok Build vs Grok 4.6: Three Products, One Name.
Frequently Asked Questions
Does Grok Bot include a model picker I can point at DeepSeek?
No. Grok Bot has no model picker for members or admins, and the teams documentation does not plan to allow admin or user choice. A Settings search for DeepSeek will not produce a serving menu. grok bot deepseek routing on this site means sending bulk summarization off the bot to a tool you already pay for, then dropping results into the workspace. Confirm DeepSeek on its own site. Do not treat failover, a Default Model line on another surface, or a pasted API key as a Bot model menu.
Can I paste a DeepSeek key into a Grok Bot charter and call that routing?
No. A Bot turn that calls an API is still Grok Bot usage on an unpublished serving set, drawing the same weekly pool, then on-demand from model and token cost. Routing means a human or an outside job writes the shorts, then a person copies those files onto the computer. Hosted MCP tokens stay with Cursor's backend and are not a DeepSeek connector you invent. Confirm any third-party API on that vendor's current page. Do not print prices from memory into the charter.
If I move two hundred PDFs off the bot this week, does overflow already billed get refunded?
No. An approval does not reverse work already completed, and pause does not unwind tokens already spent. Overflow already billed stays on the invoice. Moving the pile off the bot stops the next pass. It does not refill an unpublished weekly allowance, and it does not create a spend cap. Copy shorts into the workspace, remove PDF globs from the charter, and keep standing clocks paused until you have decided whether overflow is allowed to continue this week.
When should a summary job stay on Grok Bot instead of leaving for a cheap API?
Stay when the job needs the computer: a weekday brief from live mail and calendar, one episode from a transcript already in hand, a browser capture with a hard stop. Leave when the input is a pile you already hold and the output is shorts another tool can write. Two hundred vendor PDFs is the arbitrary pile this article uses. One podcast with timestamps can stay. A batch of dozens is a pile again. If policy forbids the second vendor, the pile waits for a human, or you accept Bot overflow with the dump still in view.