Grok 4.6 Extra High Fast in Cursor Is the Subagent Default. Pin It.

I daily-drive Cursor. The implementer hop is Grok 4.6 Extra High Fast. The Task allowlist slug is cursor-grok-4.6-xhigh-fast. I set that as the default on August 16, 2026, and every Task still has to name it — omit model and the hop inherits the parent. If the parent is a premium Claude or GPT route, that inheritance is a money leak, not a convenience. The slug stays inside Cursor. It is not an API ID.

I'm William Spurlock — founder, AI Systems Architect, and Fractional AI CTO. I've built 600+ automations with 500+ still live, spent 20,000+ hours on agentic systems, helped clients delete 35,000+ hours of busywork, and shipped hundreds of production sites. I pay the Cursor bill and the xAI invoice as two different surfaces. I do not flatten them.

Today is September 3, 2026. xAI's Grok 4.6 launch post is three weeks old. The developer page still lists one public model name: grok-4.6. Cursor has 4.6 on every plan. Extra High Fast is how I dispatch the disposable worker. If you want the editor-plus-terminal rhythm this hop sits inside, I already wrote the Cursor and Claude Code daily workflow. That post is the day shape. This one is the Task pin.

The API pin, the 200k whole-request double, and the missing Fast API ID live in Grok 4.6: Pin the ID, Watch the 200k Cliff. This post does not own that lane. Grok Bot is a standing teammate product, not this Task slug.

---

What is Grok 4.6 Extra High Fast in Cursor? #

Grok 4.6 Extra High Fast is Anysphere's Task/subagent route for Grok 4.6, not SpaceXAI's public model name. I say Extra High Fast out loud. I type cursor-grok-4.6-xhigh-fast on the Task. I type grok-4.6 in the picker and on api.x.ai. Those are three objects. One noun ("Grok") will wreck the invoice.

The developer page says Grok 4.6 is available in Cursor on all plans. The launch post said the same on August 12. Extra High Fast is the hop I added on August 16 so a premium parent can stay on the routing decision while the implementer does the file work.

Object	Spoken name	String I type	Surface
Flagship model	Grok 4.6	grok-4.6	Cursor picker, Grok Build, api.x.ai
Task default	Grok 4.6 Extra High Fast	cursor-grok-4.6-xhigh-fast	Cursor Task / subagent only
Slower 4.6 hop	Grok 4.6 Extra High	cursor-grok-4.6-xhigh	Cursor Task, only when I ask
Cheap floor	Composer 2.5 Fast	composer-2.5-fast	Shell, search, summarization
Teammate product	Grok Bot	A named bot in the Grok Bot app	Not a Task slug

I do not treat Extra High Fast as "Grok, but quicker." The Fast in the spoken name is Cursor's Task variant. The launch post mentions a fast API variant at twice the price and does not publish a public code. I will not invent grok-4.6-fast because the Task slug has -fast on the end. That hyphen folklore is how an n8n node dies on a Saturday. The cliff post owns the no-ID rule. I just refuse to import it into Cursor.

What Extra High Fast is for, on my desk:

1. Implementation hops. Multi-file edits, captions, batch writes, one-job authoring.
2. Vision-capable batch work when the Task is looking at images and writing files.
3. Keeping the parent expensive on purpose. The parent decides. The hop executes.

What it is not:

Temptation	Reality
"The cheap default if I omit model"	Omit model and you inherit the parent. There is no silent cheap fallback
"The xAI Fast ID"	There is no official Fast API ID on the developer page
"The picker string"	The picker gets grok-4.6
"A standing teammate"	Wrong object for a Task slug

If a contractor writes "we upgraded Grok in Cursor" and cannot tell me which of those four strings left the building, they did not pin anything. They renamed a vibe.

---

Why must every Task set model? #

Omitting model does not pick Extra High Fast. It inherits the parent. If the parent is a premium Claude or GPT route, the subagent is that premium route. I have burned real money on that miss. It is the first operator rule, not a style note.

The Task contract is blunt: no model field means same model as the parent agent. My August 16 default is a standing instruction for this shop. The tool does not read my mind. If a hop ships without the field, Cursor did what the contract says.

How the Task is launched	What actually runs	What I do
model: "cursor-grok-4.6-xhigh-fast"	Extra High Fast	Correct default hop
model omitted, parent is premium Claude / GPT	The premium parent	Kill it. Relaunch with the slug
model omitted, parent is already Extra High Fast	Extra High Fast by accident	Still a miss. The next parent will not be cheap
model: "cursor-grok-4.6-high-fast"	Invalid. Not on the allowlist	Do not pass it
model: "cursor-grok-4.5-high-fast"	Retired 4.5 default	Do not pass it
model: "cursor-grok-4.6-xhigh"	Slower non-fast 4.6	Only when I ask
model: "composer-2.5-fast"	Cheap floor	Shell / search / summarization

The leak is not theoretical. A parent on a thinking-heavy Claude or GPT slug that fans out ten implementer hops without model is ten premium copies. Extra High Fast exists so that fan-out is cheap enough to run and sharp enough to ship. Inheritance deletes the reason I set a default.

Pre-launch check I run in my head before every Task:

1. Is model present on this call?
2. Is the value on the allowlist I actually use?
3. If I am launching more than one hop in the same turn, does each call set model on its own?
4. If I am resuming, did I leave model off so the prior slug stays, and did I confirm that prior slug was allowed?

Resume is the other half of the same rule. A resume keeps the prior slug. Passing model on resume is the wrong move. If the prior hop was a forbidden or retired slug, I do not resume it. I launch fresh with cursor-grok-4.6-xhigh-fast.

Parallel is not a shared setting. One message, three Tasks, three model fields. A hop that "should just follow the others" is the omitted-field bug with extra steps.

Pattern	Verdict
One Task, model set to Extra High Fast	Ship
Three Tasks in one turn, each with model set	Ship
Three Tasks in one turn, only the first has model	Two inherit the parent. Kill the two
Resume with no model after a good Extra High Fast hop	Correct
Resume with a new model "just to be safe"	Wrong. Prior slug is already kept
Resume of a retired 4.5 slug	Do not resume. Launch fresh

I keep the premium parent for the routing decision. I do not hop a Task onto a Claude Opus thinking slug or a GPT flagship slug "because this one is important." Importance is a briefing problem. It is not a reason to copy the parent onto the implementer.

---

What slug do I pin — and which ones fail? #

I pin cursor-grok-4.6-xhigh-fast. I do not pass cursor-grok-4.6-high-fast. I do not pass retired cursor-grok-4.5-high-fast. The first missing word (extra / x) is how people invent a slug that looks right and is not on the allowlist. The 4.5 string is last year's default wearing this year's noun.

The allowlist I actually dispatch from Cursor:

Slug	Role	When I use it
cursor-grok-4.6-xhigh-fast	Default Task / subagent	Implementation, captions, multi-file edits, batch writes
cursor-grok-4.6-xhigh	Slower non-fast 4.6	Only when I ask for the slower 4.6
composer-2.5-fast	Floor	Shell, search, file-find, summarization, hops where reasoning depth does not matter
gpt-5.4-nano-medium	Cheapest GPT hop	Short extraction when I explicitly want GPT
gpt-5.4-mini-medium	Cost-efficient GPT hop	Longer GPT-shaped generation when I ask
gemini-3-flash / gemini-3.5-flash	Gemini hops	Vision or long-context reads when I prefer Gemini for that job

The two Grok near-misses I reject on sight:

String	Why it fails
cursor-grok-4.6-high-fast	Not on the allowlist. Looks like Extra High Fast with the x dropped
cursor-grok-4.5-high-fast	Retired default. 4.5 is not the hop
grok-4.6-fast	Not a Cursor slug and not an API ID
grok-4.6 on a Task that should be Extra High Fast	That is the picker / API ID. Wrong object
cursor-grok-4.6-xhigh-fast on api.x.ai	Cursor slug in an xAI field. Send it back

cursor-grok-4.6-xhigh is real. It is the slower 4.6. I do not "upgrade" a hop to non-fast because the job feels serious. Serious jobs get a tighter brief or they stay on the premium parent. Non-fast is a latency and depth trade I make on purpose, out loud, in the request.

composer-2.5-fast is the floor, not a second default. If the hop is ls, rg, "summarize this log," or "find the file that owns X," Extra High Fast is wasted depth. If the hop is a 12-file edit with a gate, Composer is the wrong brain.

I do not treat the allowlist as a tasting menu. Extra High Fast is the default. Everything else is an override I can point to.

---

When does Extra High Fast win, and when do I keep it off the hop? #

Extra High Fast wins implementation hops, captions, multi-file edits, and batch writes. It loses the routing decision, the trivial shell job, and the xAI token route. I do not run one slug on every surface and call that a stack.

Job	What I run	Why
Decide which files move, which model owns the week, whether the hop is even a Task	Premium parent	Routing is the expensive thought. I keep it
Implement the files the parent already scoped	cursor-grok-4.6-xhigh-fast	This is the default I set on August 16
Captions, batch markdown, multi-file writes with a gate	cursor-grok-4.6-xhigh-fast	One job, disjoint ownership, fast enough to fan out
rg, path lookup, "summarize these five lines," mechanical file move	composer-2.5-fast	Reasoning depth is irrelevant
Token call outside Cursor — n8n, MCP, a gateway	grok-4.6	Official ID on the developer page
Long API loop that might cross 200k prompt tokens	grok-4.6 plus the 200k rule	Owned by the cliff post
Slower 4.6 hop I asked for by name	cursor-grok-4.6-xhigh	Override, not a mood

The parent stays expensive because the parent is doing the part I will not delegate: what to build, what not to touch, which vendor noun is current. Extra High Fast is the hop that should not re-litigate the board. If I catch a Task rewriting the model table I pasted, the brief failed, not the slug.

I already compared editors in the complete AI coding assistant showdown. That post still describes Cursor versus the other desks. This post pins one Task slug inside Cursor. The May snapshot in the frontier comparison is history. September 3 has a current-model table further down. Paste that table into the hop. Do not let a cheap model invent last year's names.

Where Extra High Fast has earned the default on my workload:

1. Multi-file implementation with a hard file list. The parent scoped it. The hop owns those paths and nothing else.
2. Caption and content batches that need a current-model table in the prompt so the writer does not emit stale vendor nouns.
3. Vision plus write — look at a cover or a screenshot, write the markdown, stop.
4. Fan-out. Six disjoint file owners in one turn, each with model set, each with a gate.

Where I refuse the default:

Signal	Move
The hop is "figure out the architecture"	Keep the premium parent. Do not hop it
The hop is one shell command and a paste-back	composer-2.5-fast
The hop will call api.x.ai	grok-4.6, never the Cursor slug
I said "use the slower 4.6"	cursor-grok-4.6-xhigh
The prompt is a 240k dump "because 4.6 has 500k"	That is an API cliff problem. See the 200k post

xAI's own benches on the launch post are why 4.6 is on the board at all — CursorBench 69.9% at High, a real step off 4.5, not a wipe of Sol or Fable. I do not promote Extra High Fast because of that table. I promote it because it is the allowlist default I can fan out without copying the parent.

---

How do I brief a Task so Extra High Fast ships? #

One job, disjoint file ownership, the current-model table pasted in, and a gate the hop must run before it talks. Extra High Fast will happily rewrite the wrong repo if the brief is a vibe. The slug is not the brief.

I daily-drive Cursor the way I wrote in the daily workflow post: parent thinks, editor implements. The Task hop is the implementer. If I skip the file list, I paid for a second architect.

The brief I actually paste:

You own ONLY these paths:
- <absolute/path/one>
- <absolute/path/two>
Do not edit any other file. Do not commit. Do not push.

One job: <one sentence>. Stop when that job is done.

Model on this Task is already cursor-grok-4.6-xhigh-fast.
Do not send that string to api.x.ai.
Do not invent grok-4.6-fast. The public API ID is grok-4.6.

Current-model table (stale names are errors):
| Vendor | Use this | Role | Do not flatten into |
|---|---|---|---|
| OpenAI | GPT-6 Astra (gpt-6-astra) | Flagship, Sep 3 | GPT-5.6 Sol / Terra / Luna cheaper stack |
| Anthropic | Claude Fable 5.1 (claude-fable-5-1) | GA Mythos-class | Not a replacement for Opus 5 |
| Anthropic | Claude Opus 5 / Sonnet 5 / Haiku 4.5 | Default complex / volume / cheap | Opus 4.8 is a Fable fallback |
| Google | Gemini 3.8 Flash (gemini-3.8-flash) | Current Flash | 3.7 = efficiency |
| xAI | Grok 4.6 (grok-4.6) | API / picker flagship | Not Grok Bot |
| Cursor | Grok 4.6 Extra High Fast (cursor-grok-4.6-xhigh-fast) | Default Task/subagent | Not an API ID |

Gate before you stop:
1. Run: <exact command>
2. Hard-ban grep must be empty
3. Report file:line for the H1 or the function you changed, plus what the gate printed

That is the whole skeleton. No SDK walkthrough. No "while you are here" refactor. If the hop needs a second job, that is a second Task with its own model and its own paths.

Disjoint ownership is not politeness. Two Extra High Fast hops that both "own the blog folder" will collide, and the cheaper slug will not save the merge. I split by file, not by theme.

A gate is a command, not a feeling. "Looks good" is how a hop ships a stale model name. I want rg empty, a frontmatter validator quiet, or a typecheck that exits 0. Extra High Fast will run the command if I write it. It will not invent a gate I was too lazy to name.

What I put in the brief versus what I keep on the parent:

Belongs in the Task	Stays on the parent
Exact paths	Whether the job should exist
One-sentence job	Vendor choice for the week
Current-model table	Whether to use Astra, Fable, or 4.6 on the API
Gate command	Whether to spend the hop at all
Hard nos (no commit, no other files, no invented Fast ID)	Client politics, budget, ship/no-ship

If the parent cannot write that table, the parent is not ready to hop. Extra High Fast is fast. It is not a substitute for a decision.

Briefing mistakes I still catch on my own desk:

Brief I almost sent	What Extra High Fast does with it	Fix
"Clean up the blog folder"	Touches every markdown file it can see	Name two paths. Stop
"Use your judgment on the model"	Invents a noun or copies the parent	Paste the September 3 table
"And also fix the API node"	Writes a Cursor slug into xAI	Second Task, or keep API work on the parent
"Make it better"	Rewrites voice, structure, and scope	One sentence job. Gate named
"Same as the last hop"	Inherits a path list I did not repeat	Paste the paths again

I do not brief a hop with "you are a senior engineer." I brief it with paths, a current-model table, and a command that can fail.

---

What do I check when the hop comes back? #

I re-run the gate myself and I spot-check the file:line claims. I do not take Extra High Fast's word that it passed. The hop is the implementer. The parent is still on the hook for the merge.

August 16 did not make me less suspicious. It made the hop cheap enough that I can afford to throw a bad one away. If the gate is green and the cited line is a different heading, I send it back. If the hop edited a third file "for consistency," I revert that file and tighten the path list.

After-hop checklist I actually use:

1. Re-run the named gate on the parent. The hop's paste can lie. The command cannot.
2. Open the cited file:line. If the H1 or the function is not where they said, the report is theater.
3. Diff the owned paths only. A surprise third file is a brief violation, not a bonus.
4. Grep the stale nouns I banned in the prompt. Cheap hops are the ones most likely to emit last year's names.
5. Confirm no Cursor slug landed in an API field. That is the other post's lane leaking into this one.

Hop report says	I believe it when
"Gate passed"	I ran the same command and saw the same exit
"Only touched the owned files"	git status / the diff matches the path list
"Did not invent a Fast ID"	Grep for grok-4.6-fast and the Cursor slug on api.x.ai paths is empty
"Used the current-model table"	The prose names Astra, Fable 5.1, 3.8 Flash, and grok-4.6 — not a flatten
"Ready to ship"	I said ship. The hop does not get that verb

Failure modes I have already paid for, so you do not have to:

Failure	What it cost me	Standing fix
Ten hops, no model, premium parent	Ten copies of the expensive route	model on every Task, including parallel
Resume of a retired 4.5 slug	A hop that looked current and was not	Launch fresh. Do not resume junk
One path list, two hops	Colliding writes in the same folder	Disjoint files or a single hop
"Also wire the xAI node" inside a Cursor Task	Cursor slug in a production HTTP field	Split the job. API gets grok-4.6
Gate described, never named	A hop that declared victory	Exact command in the brief

The daily workflow post still owns morning planning versus in-editor flow. This checklist is the missing piece that post could not have: Extra High Fast did not exist as my default when I wrote it. September 3 is the pin. The day shape did not change. The hop string did.

If you run a wider agentic layer around Cursor, the OS still owns budget and approval. I already said that in what an agentic OS means. Extra High Fast does not get a send button. It gets a path list and a gate.

---

Why can't I send Extra High Fast to api.x.ai? #

Because cursor-grok-4.6-xhigh-fast is a Cursor Task slug. The public API ID is grok-4.6. The developer page prints one model name. I will not type Anysphere's hop into SpaceXAI's field.

This is the only API paragraph I need here. The 200k whole-request double, cache keys, reasoning_effort, and the official "there is no Fast ID" rule live in Grok 4.6: Pin the ID, Watch the 200k Cliff. If you want the invoice math, read that. If you want the Cursor hop, stay on this page.

String	Valid on a Cursor Task?	Valid on api.x.ai?
cursor-grok-4.6-xhigh-fast	Yes. Default	No
cursor-grok-4.6-xhigh	Yes, when I ask	No
grok-4.6	Wrong object for a Task hop	Yes. Only official ID
grok-4.6-fast	No	No. Not on the developer page

The launch post says a fast variant exists at twice the price. A sentence is not an ID. Extra High Fast is how Cursor sells me speed on the hop. It is not a hint that I should hyphenate grok-4.6.

If a pull request puts cursor-grok-4.6-xhigh-fast in an HTTP node, I send it back. If a Task tries to "help" by writing that slug into an env file for xAI, I send that back too. Two surfaces. Two strings.

---

Where does Extra High Fast sit on the September 3 board? #

It is the Cursor Task default. It is not the OpenAI flagship, not Anthropic's Mythos-class GA, not Google's current Flash, and not the xAI API pin. Today GPT-6 Astra shipped. I already wrote the Astra operator spec card. I am not converting the studio to one vendor because xAI has a good hop inside Cursor.

The board I paste into Tasks today:

Vendor	Use this	Role	Do not flatten into
OpenAI	GPT-6 Astra (gpt-6-astra)	Flagship, Sep 3	GPT-5.6 Sol / Terra / Luna cheaper stack
Anthropic	Claude Fable 5.1 (claude-fable-5-1)	GA Mythos-class	Not a replacement for Opus 5
Anthropic	Claude Opus 5 / Sonnet 5 / Haiku 4.5	Default complex / volume / cheap	Opus 4.8 is a Fable fallback
Google	Gemini 3.8 Flash (gemini-3.8-flash)	Current Flash	3.7 = efficiency
xAI	Grok 4.6 (grok-4.6)	API / picker flagship	Not Grok Bot
Cursor	Grok 4.6 Extra High Fast (cursor-grok-4.6-xhigh-fast)	Default Task/subagent	Not an API ID

Last spring's three-vendor map is the May 2026 frontier post. It is a snapshot. It is not this table. A hop that still writes last year's names failed the paste.

How I run the board inside Cursor today:

Surface	String	Notes
Parent / routing chat	Premium Claude or GPT, on purpose	Decides. Does not implement the folder
Editor picker for a 4.6 pass I will watch	grok-4.6	Official ID, all Cursor plans
Task implementer	cursor-grok-4.6-xhigh-fast	Default since August 16
Trivial hop	composer-2.5-fast	Floor
Outside Cursor	grok-4.6	Plus the 200k rule on the other post

The rest of the day still sits in an agentic OS I already described in what an agentic OS means for running a business. Extra High Fast is a worker in that OS. It is not the OS.

---

What will I not claim about Extra High Fast on September 3? #

I will not publish a Cursor price I did not measure, a Fast API ID xAI did not print, or a bake-off I did not run this morning. A pin is a pin. It is not a keynote.

Claim I will not make	Why
"Extra High Fast is $X per hop"	I am not inventing Cursor pricing
"grok-4.6-fast is the latency ID"	Not on the developer page
"Omit model and you get the cheap default"	You inherit the parent
"cursor-grok-4.6-high-fast is close enough"	Not on the allowlist
"The August 12 first-week 2x Cursor promo is still on"	Launch-week offer. Ship date was August 12
"Extra High Fast beats Astra / Fable / 3.8 Flash in my shop"	I did not run that bake-off today
"This slug replaces Grok Bot"	Different product. One sentence. Already said

I will claim the pin: every Task sets model, the default is cursor-grok-4.6-xhigh-fast as of August 16, near-miss slugs are errors, resume keeps the prior slug, parallel hops each name their own, and the API still gets grok-4.6.

---

Frequently asked questions #

How should operators use Grok 4.6 Extra High Fast in Cursor? #

Pin cursor-grok-4.6-xhigh-fast on every substantive Task, keep the premium parent for the routing decision, and use composer-2.5-fast when the hop is shell or search. Brief one job, give disjoint paths, paste the current-model table, demand a gate. Never send the Task slug to api.x.ai. The API lane is the 200k cliff post.

What is the Cursor Task slug for Grok 4.6 Extra High Fast? #

cursor-grok-4.6-xhigh-fast. That is the allowlist string I set as the default on August 16, 2026. The spoken name is Grok 4.6 Extra High Fast. The picker and the API still use grok-4.6. Do not drop the x. Do not write high-fast and hope.

What happens if I omit the model parameter on a Cursor Task? #

The subagent inherits the parent. If the parent is a premium Claude or GPT route, the hop is that premium route. There is no silent Extra High Fast fallback. I treat a missing model as a kill-and-relaunch, not a warning.

Is cursor-grok-4.6-high-fast a valid slug? #

No. It is not on the allowlist. The valid Fast hop is cursor-grok-4.6-xhigh-fast. The retired string cursor-grok-4.5-high-fast is also invalid. Near-miss slugs are how a hop fails before it writes a file.

When should I use composer-2.5-fast instead? #

When the hop is shell, search, a mechanical file move, or a short summary and reasoning depth does not matter. Extra High Fast is the default for implementation and batch writes. Composer is the floor. I do not run Extra High Fast at ls.

Should I pass model when I resume a Task? #

No. Resume keeps the prior slug. Confirm that prior slug was allowed. If it was a retired or forbidden value, do not resume — launch a new Task with cursor-grok-4.6-xhigh-fast. Passing model on resume is the wrong shape.

Can I send cursor-grok-4.6-xhigh-fast to the xAI API? #

No. That is a Cursor Task slug. The developer page lists grok-4.6 as the model name. The launch post mentions a fast variant and does not give a public code. For the 200k cliff and the no-Fast-ID rule, read the cliff post.

Is Extra High Fast the same as Grok Bot? #

No. Grok Bot is a standing teammate product, not this Task slug. Extra High Fast is a disposable Cursor worker. I do not staff a bot by typing an allowlist string, and I do not dispatch a Task by opening the bot app.

When does Extra High Fast win versus a premium parent? #

It wins the implementer hop — files, captions, multi-file edits, batch writes. The premium parent wins the routing decision: what to build, which vendor is current, whether to hop at all. If the Task is still deciding the architecture, I kept the wrong brain on the hop.

Do parallel Tasks share one model setting? #

No. Each Task in the same turn needs its own model field. A hop that "follows the others" inherits the parent. Three implementers means three Extra High Fast pins, three path lists, three gates.

What is cursor-grok-4.6-xhigh used for? #

The slower non-fast 4.6 hop, and only when I ask for it. It is on the allowlist. It is not the default. I do not "promote" a job to non-fast because it feels important. Importance stays on the parent or in the brief.

Did I set Extra High Fast as the default on August 16? #

Yes. August 16, 2026. The standing instruction in this shop is: every substantive Task sets model to cursor-grok-4.6-xhigh-fast unless I named a different allowlist slug. The date is the receipt. The field on the Task is the enforcement.

Does Extra High Fast replace GPT-6 Astra or Claude Fable 5.1? #

No. Astra is today's OpenAI flagship — see the spec card. Fable 5.1 is Anthropic's GA Mythos-class model and is not an Opus replacement. Extra High Fast is the Cursor implementer. I pin it on Tasks. I do not flatten September into one hop.

---

If omitted-model hops are still copying a premium parent, that is Fractional AI CTO work, not a Grok upgrade. Book an AI automation strategy call and I will pin cursor-grok-4.6-xhigh-fast on the implementer Tasks, drop shell hops to composer-2.5-fast, and keep that slug out of api.x.ai — or we stand up a custom agent team with those gates already named. I have done this across 600+ automations. The leak is almost never "we need more Grok." It is a missing model field on a hop that inherited Claude.