Grok 4.6: Pin the ID, Watch the 200k Cliff.

I missed pinning Grok 4.6 on August 12. Today I pin grok-4.6, I treat 200,000 prompt tokens as a cliff that doubles the whole request, I keep xhigh on 4.6 only, and I will not invent a grok-4.6-fast API ID. This is the API and picker pin. If you want the Cursor Task default, I already wrote Grok 4.6 Extra High Fast in Cursor.

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 xAI invoice. I do not collect launch adjectives.

Today is September 4, 2026. The Grok 4.6 launch post is three weeks old. The developer page and the model card are the pin documents. If you want last spring's three-vendor snapshot, I already wrote the May 2026 frontier comparison. That post is history. This one is the catch-up I am running in Cursor and on api.x.ai.

---

What did I miss when Grok 4.6 shipped on August 12? #

I missed a model pin, not a new product. xAI shipped Grok 4.6 on August 12 as the flagship brain for coding, long agent loops, and knowledge work — available that day in Cursor, Grok Build, the xAI API, and the usual gateways. I kept spending on older Grok routes and Cursor Task slugs while the public ID sat in the docs. That is the miss.

The launch post is blunt about the job: stay with a complex task across many steps, whether that is research, a codebase, or turning a product idea into a first working artifact. It also published a first-week 2x usage promo inside Grok Build and Cursor. That week is over. I am not writing this as if the promo is still live.

Here is the ship card I should have pinned on day one:

Fact	Value	Official source
Ship date	August 12, 2026	Launch post
API ID	grok-4.6	Developer page
Context	500,000 tokens	Same
Knowledge cutoff	January 2026	Same
Inputs / output	Text and image in; text out	Same
Output cap	No text output limit	Same
Headline price	$2 input / $6 output per 1M	Launch post and developer page
Cached input under 200k	$0.50 per 1M	Model card and release notes
Reasoning	low, medium, high (default), xhigh	Developer page + reasoning guide
Surfaces	xAI API, Grok Build, Cursor (all plans), OpenRouter, Vercel, Cloudflare	Developer page

The August release notes repeat the same ID, the 500k window, the 200k price bands, and the four effort levels. If a contractor still has grok-4.5 hard-coded "because 4.6 is just a Cursor thing," they did not read the notes.

xAI's published benches, from the launch post, are why I bother to catch up. These are their numbers, not a bake-off I ran this morning:

Eval (xAI table, Grok 4.6 High)	Grok 4.6 High	Grok 4.5 High	GPT-5.6 Sol Max	Fable 5 Max
AA Intelligence Index	61	56	61	62
GDPVal-AA v2	1753	1526	1728	1741
CursorBench v3.2	69.9%	66.7%	67.2%	70.5%
DeepSWE v1.1	65.9%	54%	73%	70%
FrontierCode v1.1 (Extended)	61.3%	56.6%	60.6%	63.6%
APEX-Agents	57.5%	47.1%	56.7%	59.2%
Terminal-Bench v3.0	26%	15.7%	34.6%	34.1%

I do not promote a client route because xAI posted a 61 on a composite index. I do notice the pattern: 4.6 is a real step off 4.5 on agent and coding benches, and it is not a wipe of Sol or Fable. Terminal-Bench still sits well behind both. That is a pin fact, not a slogan.

What I missed in practice:

1. The public API ID. grok-4.6 has been callable since August 12. I left nodes on 4.5.
2. The 200k rule. The headline $2 / $6 is the short-context band. Cross 200k and the whole request moves.
3. xhigh. That effort exists on 4.6. It does not exist as a real setting on 4.5.
4. The Fast temptation. The launch post mentions a fast variant at twice the price and does not hand me a public ID I will type into n8n.

If you want the daily Cursor / Claude Code rhythm this pin sits inside, I already wrote the Cursor and Claude Code daily workflow. This post does not replace that. It replaces a stale Grok string.

---

What model ID do I pin in Cursor and the API? #

The ID is grok-4.6. That is the string on the developer page, the model card, and the release notes. I pin that string in the xAI API, in Grok Build, and in any gateway that exposes the same code. I do not pin a Cursor Task slug into api.x.ai.

The developer page's own "at a glance" table is the card I keep next to the invoice:

Spec	Value I pin	Source
Model name	grok-4.6	Developer page
Context window	500,000 tokens	Same
Knowledge cutoff	January 2026	Same
Modalities	Text and image input; text output	Same
Output limit	None on text	Same
APIs	Responses API and Chat Completions	Same
Tools	Function calling, web search, X search, code execution	Same
Batch API	Not supported	Model card
Rate limits	150 requests/sec, 50,000,000 tokens/min	Same
Regions	us-east-1, us-west-2	Same

Where it runs, from the same developer page:

Surface	What I do
xAI API	Pin grok-4.6. Set prompt_cache_key on Responses, or x-grok-conv-id on Chat Completions
Grok Build	Accept it as the default coding-agent model. Do not invent a second Build slug
Cursor	Use 4.6 in the editor. Treat cursor-grok-4.6-xhigh-fast as a Task slug, not an API ID
OpenRouter / Vercel / Cloudflare	Only if the gateway's listed ID is still grok-4.6. I do not accept a "fast" alias from a reseller page

Cursor is not a side channel. The developer page says Grok 4.6 is available on all Cursor plans. The launch post said the same on day one. If your team "doesn't have Grok in Cursor," that is a picker problem, not a plan problem.

The three objects I refuse to flatten:

Object	What it is	What I type
Grok 4.6	Flagship model	grok-4.6
Grok Bot	Always-on teammate product, shipped August 11	A named bot in the Grok Bot app, not a model ID
Cursor Task slug	Disposable editor worker	cursor-grok-4.6-xhigh-fast inside a Cursor Task. Never on api.x.ai

If a pull request puts cursor-grok-4.6-xhigh-fast in an n8n HTTP node, I send it back. That slug is Anysphere's Task route. It is not SpaceXAI's model name.

I also do not write "Grok" next to a raw 500k and call it a spec. I write the ID, then the integer, then the price band. The May frontier post already taught me what happens when a studio talks in vendor nouns instead of IDs. You get a week of "we upgraded Grok" with nobody able to say which string left the building.

---

What is the Grok 4.6 200k billing cliff? #

Once the prompt reaches 200,000 tokens, xAI bills the higher rate for every token in that request — input, cached input, and output — not just the slice above 200k. That sentence is on the Grok 4.6 model card: "Requests whose prompt reaches 200k tokens are billed at the higher rate for all tokens in the request." The release notes print the same bands.

The official price table:

Type	Prompt under 200k (per 1M)	Prompt at or above 200k (per 1M)
Input	$2.00	$4.00
Cached input	$0.50	$1.00
Output	$6.00	$12.00

The developer page still leads with $2 / $6. That is not a contradiction. That is the short-context sticker. If you quote $2 / $6 on a 300k agent loop, you are quoting the wrong band.

The cliff is a step, not a ramp. One extra token that pushes the prompt from 199,999 to 200,000 does not add a surcharge on that token. It re-prices the whole request.

Operator math I actually use — prompt tokens only, no cache, using the published rates:

Prompt tokens	Input band	Input cost	What changed
100,000	$2.00 / 1M	$0.20	Safe short-context band
180,000	$2.00 / 1M	$0.36	Still under the line
199,000	$2.00 / 1M	$0.40	Last cheap band I will plan for
200,000	$4.00 / 1M	$0.80	Same work, 2x input on the whole prompt
350,000	$4.00 / 1M	$1.40	You bought the 500k window. You also bought the high band

Output follows the same switch. A 10,000-token answer is $0.06 under the line and $0.12 over it. Cached input is $0.50 / $1.00. If you are looping a long agent, you pay the high band on the prompt, the high band on the cache read, and the high band on the tokens the model writes. That is why I say the cliff doubles the whole request.

What I do with that rule:

Prompt size I see	Move
Under 120k	Leave it. Do not compact for sport
120k–180k	Watch the next tool dump. Compaction is cheaper than a surprise 200k
180k–199k	Stop appending files "for context." You are one paste from the high band
200k+	Only if I meant to buy the high band for that job. Log it. Do not discover it on the invoice

The 500,000-token window is not a dare. It is a ceiling. I already wrote how I calculate automation ROI before I build. A 4.6 loop that lives at 280k because nobody compacted is not "using the context." It is paying 4.5-era work at the long-context sticker.

I do not treat 199k as a target. I treat 180k as the last comfortable working set. The extra 19k is buffer for one more tool result, not a place to store a second repo.

---

How should I set reasoning_effort, including xhigh? #

Default is high. I leave most API calls there. I use low on latency-first tool loops, medium when the job is long-context analysis that is not a hard proof, and xhigh only on Grok 4.6 for a pass I will sit with. The reasoning guide says effort cannot be turned off on 4.6 or 4.5. If you omit the field, you get high.

Official effort table, from that guide:

reasoning_effort	xAI's description	Where I set it
low	Some reasoning tokens, still fast	Latency-sensitive agents, simple tool calls, routers
medium	More thinking, less latency-sensitive	Long-context analysis I will not watch live
high (default)	Deeper thinking, more tokens	Default grok-4.6 pin
xhigh	Maximum depth, higher latency	One hard planning pass on 4.6. Never a classifier

The same guide is the xhigh rule I will not let a contractor blur:

"xhigh" is available on grok-4.6 and later. On models that do not support it, such as grok-4.5, requests with "xhigh" are treated as "high".

That is the whole operator lesson. xhigh is a 4.6 feature. Send it to 4.5 and you paid for a keyword that silently became high. Your logs will say xhigh. The model will not.

Model	Effort values that actually change behavior	What happens if I send xhigh
grok-4.6	low / medium / high / xhigh	Maximum depth. You pay the thinking tokens
grok-4.5	low / medium / high	Treated as high. No extra depth
grok-4.20-multi-agent	Same four strings, different meaning	Controls agent count, not depth. I do not pretend that is 4.6

I do not send presencePenalty, frequencyPenalty, or stop on these calls. The reasoning guide says those fields error on reasoning models. Dead knobs in a payload make logs lie. Same class of mistake as leaving thinking_budget on a Gemini 3.8 Flash route.

Reasoning tokens are not a side meter. The guide says they are billed as part of total consumption. xhigh does not change the $2 / $6 or $4 / $12 sticker. It changes how many tokens you mint before the first sentence you will read. Stack that on a 210k prompt and you have the cliff and a long think on the same invoice.

My pin for September 4:

Route	Model	Effort
Default API / n8n Grok node	grok-4.6	high
Cheap router / "which folder is this"	grok-4.6	low
Long research I will not watch	grok-4.6	medium, and stay under 180k
Architecture brief I will read	grok-4.6	xhigh, one shot, then drop back
Anything still on 4.5	grok-4.5	high at most. No xhigh theater

If the job is "always-on teammate," that is Grok Bot, not an xhigh flag. If the job is a Cursor implementer hop, pin Extra High Fast.

---

Is there a grok-4.6-fast ID I can pin? #

No. The developer page lists one model name: grok-4.6. I will not invent grok-4.6-fast. The launch post says there is a fast variant at twice the price and does not publish a public code I will type into production.

That is the whole Fast rule. A sentence in a launch post is not an ID. A Cursor Task slug is not an ID. A reseller dropdown labeled "Grok 4.6 Fast" is not an ID until SpaceXAI prints it on the developer page the same way they printed grok-4.6.

String someone will try	What it actually is	What I do
grok-4.6	Official API ID	Pin it
grok-4.6-fast	Not on the developer page	Reject the PR
grok-4.6-latest	Not a documented 4.6 alias	Reject it
cursor-grok-4.6-xhigh-fast	Cursor Task slug	Keep it inside Cursor Tasks
Older grok-4-fast-reasoning / grok-4-1-fast-reasoning	Retired 4.x Fast slugs that redirect toward grok-4.3	Do not "upgrade" those to a made-up 4.6 Fast

The May 15 retirement note is why people still have Fast in their muscle memory. Those old Fast IDs were real. They are not a template for 4.6. If a node still says grok-4-fast-reasoning, the fix is a documented current ID — grok-4.6 with an explicit reasoning_effort, or grok-4.3 if you still want that cheaper, faster stack — not a hyphen you invented because the launch post said "fast variant."

Latency I actually control on 4.6:

1. Set reasoning_effort to low. That is the documented fast-ish setting on this model.
2. Keep the prompt under 180k. The cliff is slower in the only way that matters: it costs more for the same tokens.
3. Set prompt_cache_key. A cache-cold 150k prompt feels "slow" and bills like a first visit.
4. Do not type a Fast ID. There is nothing to pin.

If a gateway shows a 4.6 Fast SKU at 2x, I treat that as their product name until I can see grok-4.6 plus a second official code on docs.x.ai. I have been burned by hyphen folklore before. I am not doing it on a model I already missed for three weeks.

---

How is Grok 4.6 different from Grok Bot? #

Grok 4.6 is the model ID grok-4.6. Grok Bot is the always-on teammate product xAI shipped on August 11. One is a brain you call. The other is a seat you staff. I use both. I do not buy a token price and expect a coworker, and I do not buy a named bot and expect a 200k price band.

xAI put them one day apart on purpose. Grok Bot launched August 11. Grok 4.6 launched August 12. People have been flattening those two nouns for three weeks. That flatten is how you ask a cloud VM for a cached-input rate, or ask grok-4.6 to text you overnight.

Axis	Grok 4.6	Grok Bot
What it is	Flagship model	Named teammate on a shared cloud computer
Ship date	August 12, 2026	August 11, 2026
How I call it	grok-4.6 in API, Cursor, Grok Build	A bot in the Grok Bot app
Billing I care about	Per-million tokens, 200k cliff, cache	Product usage, separate from Cursor and Grok plan tokens
Effort knob	reasoning_effort including xhigh	Not a model card
Memory	Whatever I put in the prompt / cache key	The thread and the computer persist when I close the lid

The Cursor implementer hop lives in Grok 4.6 Extra High Fast in Cursor. This post does not restaff a bot roster. If you want the broader definition of agentic software, start with what agentic AI is in 2026. If you want the day-to-day OS, read what an agentic OS means.

The only sentence I need here: pin the model in the editor and the API; staff the bot in the bot app.

---

How do I keep cache hits and stay off the cliff? #

Set a prompt_cache_key on the Responses API (or x-grok-conv-id on Chat Completions), only append to the thread, and compact before you cross 180k. The developer page says that without the key you often pay full input on a cache-cold server. The caching best-practices page makes the key the first rule.

Cached input on 4.6 is $0.50 / 1M under 200k and $1.00 / 1M at or above 200k, per the model card. That is still a discount. It is not a reason to drag a 240k prompt because "cache will save it." Cache at the high band is $1.00, not $0.50. The cliff hits the cache line too.

What the What Breaks Caching page will punish:

Mistake	Result	Operator fix
No prompt_cache_key / no x-grok-conv-id	Requests land on a cold server	Set a stable conversation ID
Edit an earlier message	Prefix match dies	Append only
Delete a turn to "clean the thread"	Prefix match dies	Compact with the API, do not hand-edit history
Reorder system / user / tool turns	Prefix match dies	Front-load static context once
Drop reasoning_content on a later turn	Top miss on reasoning models	Keep the reasoning payload the docs ask for

The developer page also points long agent loops at context compaction. That is the other half of the cliff. Compaction is how I stay in the $2 / $0.50 / $6 band without pretending a 500k window is free.

My loop hygiene on 4.6:

1. One cache key per job, not per HTTP call. A UUID or the workflow run ID is enough.
2. Static rules and schemas at the front. If the system prompt changes every turn, you paid for a new prefix.
3. Tool results get summarized before they age. A 40k log dump is how you walk from 160k to 210k in one node.
4. Watch cached_tokens. The best-practices page says if that number stays 0, the key or the order is wrong.
5. Do not compact for sport under 120k. Compaction is a cliff tool, not a style choice.

If you run this inside a standing ops layer, the OS still owns the budget. I already said that in the agentic OS post. A 4.6 node that appends forever is not an agent. It is an unbounded invoice.

---

Where does Grok 4.6 sit on the September 2026 board? #

Grok 4.6 is xAI's flagship model. It is not GPT-6 Astra, not Claude Fable 5.1, not Gemini 3.8 Flash, and not Grok Bot. I catch up the pin. I do not convert the studio to one vendor because I was three weeks late.

The board I am running this week:

Vendor	Use this	Role	Do not flatten into
OpenAI	GPT-6 Astra (gpt-6-astra)	Flagship, September 3	GPT-5.6 Sol / Terra / Luna stay the cheaper stack
Anthropic	Claude Fable 5.1 (claude-fable-5-1)	GA Mythos-class, September 1	Not a replacement for Opus
Anthropic	Claude Mythos 5.1 (claude-mythos-5-1)	Same weights, invite-only	Not a public default
Anthropic	Claude Opus 5 / Sonnet 5 / Haiku 4.5	Default complex / volume / cheap	Opus 4.8 is a Fable fallback, not the flagship
Google	Gemini 3.8 Flash (gemini-3.8-flash)	Current Flash, September 2	3.7 Flash = efficiency fallback; 3.1 Pro = preview
xAI	Grok 4.6 (grok-4.6)	Flagship model, August 12	Not Grok Bot
xAI	Grok Bot (Mac app 0.43.0)	Always-on agent product, August 11	Not Cursor cursor-grok-4.6-xhigh-fast

Where 4.6 goes on my stack after this pin:

Job	Model I call	Why
Cursor coding pass I will read	grok-4.6 at high	Launch pitch is long-running agents and first-pass apps
Hard planning brief	grok-4.6 at xhigh, once	4.6-only effort. Then drop back
Cheap classification	Not 4.6 at xhigh	low on 4.6, or a cheaper ID
Always-on ops that must survive a closed laptop	Grok Bot	Product, not a token ID
OpenAI flagship / computer-use fence	GPT-6 Astra	Different vendor, different card
Anthropic GA Mythos-class	Claude Fable 5.1	Not an Opus replacement
Google volume Flash	Gemini 3.8 Flash	Keep 3.7 on the cheap loop

I already compared coding assistants in the complete AI coding assistant showdown. That post still describes the editors. This post pins one model inside Cursor. If you are looking for Antigravity recipes, those live in the Antigravity agents blueprint and Antigravity 2 subagent recipes. I do not route those through a Grok ID.

---

What Cursor prompt do I use to pin this stack? #

I paste a pin, not a rewrite. The job is to change strings, effort, and cache keys — not to invent a Fast ID or merge Grok Bot into the API.

Catch-up pin for Grok 4.6. Today is September 4, 2026. Do not invent architecture.

For every xAI / Grok route that should move:
1. Set model to grok-4.6. That is the only official ID.
2. Do not add grok-4.6-fast, grok-4.6-latest, or any Fast hyphen. The launch post mentions a fast variant and does not give a public code.
3. Do not send cursor-grok-4.6-xhigh-fast to api.x.ai. That slug stays inside Cursor Tasks.
4. Set reasoning_effort to high unless I marked the route as a router (low) or a single planning pass I will read (xhigh).
5. Never send xhigh to grok-4.5. On 4.5 that string becomes high.
6. Set prompt_cache_key on Responses (or x-grok-conv-id on Chat Completions). One key per job.
7. Keep prompts under 180k unless I explicitly accept the 200k whole-request double ($4 / $1 / $12).
8. Do not flatten this into Grok Bot. Bot staffing is a different product.

Then print a table: route name, old ID, new ID, effort, cache key present, last measured prompt tokens, cliff risk.

That is the whole instruction. No SDK walkthrough. No "while we are here" refactor of the n8n graph.

If a route is still on a retired Fast slug from the May 15 note, the same prompt applies: pick a current documented ID. Do not "port" Fast by adding -fast to 4.6.

I daily-drive Cursor. The picker gets grok-4.6. The Task dispatcher can keep cursor-grok-4.6-xhigh-fast for one-pass editor work. Those are two surfaces. One pin document.

---

What will I not claim about the 200k cliff on September 4? #

I will not claim a bake-off I did not run, a Fast ID xAI did not publish, or a first-week promo that expired in August. Catch-up is a pin. It is not a new benchmark blog.

Hard no list:

Claim I will not make	Why
"4.6 beats Astra / Fable / 3.8 Flash in my shop"	I did not publish a measured bake-off today
"grok-4.6-fast is the latency ID"	Not on the developer page
"The Cursor 2x promo is still on"	Launch-week offer. Ship date was August 12
"Grok Bot is just 4.6 with a nicer UI"	Different product. Different bill
"xhigh on 4.5 is fine"	Docs say it becomes high
"500k means dump the repo"	200k re-prices the whole request
"I independently verified the launch benches"	Those numbers are xAI's table

I will claim the pin: grok-4.6, 500k window, $2 / $6 under 200k, whole-request double at 200k, xhigh on 4.6 only, cache key required, Fast hyphen rejected.

If a client asks "is this the new best model," the answer is "it is xAI's current flagship, and I am three weeks late pinning it." That sentence is cheaper than a weekend of hyphen folklore.

---

Frequently asked questions #

How should operators use Grok 4.6 in Cursor and the API? #

Pin grok-4.6, default reasoning_effort to high, set a cache key, and stay under 180k prompt tokens unless you mean to pay the 200k whole-request double. In Cursor, use 4.6 in the picker and keep cursor-grok-4.6-xhigh-fast inside Tasks. On api.x.ai, type the official ID only. That split is the whole operator move.

What is the Grok 4.6 model ID? #

grok-4.6 — that is the only model name on the developer page, the model card, and the August release notes. Do not send a Fast suffix. Do not send a Cursor Task slug. Pin the string.

What happens when a Grok 4.6 prompt hits 200k tokens? #

The whole request moves to $4 input / $1 cached input / $12 output per million tokens. The model card says the higher rate applies to all tokens in the request, not the slice above 200k. A 199k prompt and a 200k prompt are different invoices, not a rounding error.

Does the 200k cliff apply only to tokens above 200k? #

No. It re-prices the entire request. Input, cached input, and output all switch bands once the prompt reaches 200k. That is why I call it a cliff and why I compact at 180k instead of surfing 199k.

Is xhigh available on Grok 4.5? #

No as a real setting. The reasoning guide says xhigh is available on grok-4.6 and later; on grok-4.5 the same string is treated as high. Your logs can still say xhigh. The model will not go deeper. Pin 4.6 if you want the extra depth.

What reasoning_effort should I default to on Grok 4.6? #

high. That is the documented default, and reasoning cannot be disabled. Use low on routers and simple tool calls. Use medium on long analysis you will not watch. Use xhigh on one planning pass you will read, then drop back. xhigh on a classifier is how a $2 / $6 card still hurts.

Is there a grok-4.6-fast API ID? #

No. The developer page does not list one. The launch post mentions a fast variant at twice the price and does not publish a public code. For latency, set reasoning_effort to low and keep the prompt short. Do not invent a hyphen.

Should I pin cursor-grok-4.6-xhigh-fast in the xAI API? #

No. That is a Cursor Task slug, not the API ID. I dispatch it for a one-pass editor worker. I send grok-4.6 to api.x.ai. Mixing those strings is how an n8n node fails on a Saturday.

How is Grok 4.6 different from Grok Bot? #

Grok 4.6 is the model. Grok Bot is the August 11 teammate product. I pin grok-4.6 on the API and picker. I pin Extra High Fast on Cursor Tasks. I staff a standing bot in the Grok Bot app. Those are three surfaces.

What is the Grok 4.6 context window and price? #

500,000 tokens. $2 input / $0.50 cached / $6 output per million under 200k prompt tokens, and $4 / $1 / $12 at or above 200k, per the model card and release notes. The developer page still leads with $2 / $6 because that is the short-context sticker. Quote the band you are actually in.

How do I keep Grok 4.6 cache hits? #

Set prompt_cache_key on Responses or x-grok-conv-id on Chat Completions, only append new turns, and watch cached_tokens. The caching best-practices page puts the key first. Editing, deleting, or reordering earlier messages breaks the prefix, per What Breaks Caching.

Where can I call Grok 4.6 today? #

xAI API, Grok Build, Cursor on all plans, plus OpenRouter, Vercel, and Cloudflare, per the developer page. Batch API is not supported on the model card. Rate limits there are 150 requests per second and 50 million tokens per minute in us-east-1 and us-west-2.

Should I dump a whole repo into the 500k window? #

No. The window is a ceiling. The bill changes at 200k. I keep working sets under 180k and compact when a tool dump would cross the line. A 500k paste is how you buy the high band for a job that fit in the cheap band yesterday.

Does Grok 4.6 replace GPT-6 Astra or Claude Fable 5.1? #

No. 4.6 is xAI's flagship, not a one-model shop. Astra is the OpenAI flagship lane. Fable 5.1 is Anthropic's GA Mythos-class model and is not an Opus replacement. Gemini 3.8 Flash is Google's current Flash. I pin 4.6 on Grok routes. I do not flatten September into one vendor.

---

If a 240k prompt is doubling the whole Grok request, or someone typed grok-4.6-fast into production, that is the catch-up I run. Book an AI automation strategy call and I will pin grok-4.6, put the 180k compact line on the loop, and keep Fast hyphens out of n8n — or we wire a custom agent team that already knows the 200k cliff. I have done this across 600+ automations. Three weeks late is cheaper than a weekend of invented IDs.