Four companies sell Grok 4.6 and all four print $2.00 and $6.00. The whole difference between them is what happens at 200,000 tokens, and Microsoft's answer is to not sell you the 300,000 tokens after it.

Twelve figures, four vendors, zero disagreement

Input, cached input, output. xAI charges $2.00, $0.50 and $6.00 per million tokens. So does Google on Vertex. So does AWS on Bedrock's Global cross-Region card. So does Microsoft on Azure Foundry, in the one table it has published for this model. Twelve numbers, and not one of them disagrees with its counterpart.

That is the table that gets copied, and taken alone it says the choice of venue is a procurement question rather than a pricing one. Buy from whoever you already have a contract with. Every comparison of these four we could find reaches that conclusion, and it is wrong by a factor of two for anybody sending long prompts, because the interesting line on this model's rate card is not on the rate card.

Bedrock's In-Region and Geo routing options are 1.10x the Global figures quoted above. Azure's are Global Standard, the only deployment type Microsoft offers here. All read August 30, 2026.

xAI sells this model with a tripwire in it

Grok 4.6 has a 500,000-token context window and a second rate card that switches on at 200,000. xAI's pricing table, last updated August 21, gives the model a row labelled long context at 200k tokens and above, where input becomes $4.00, cached input $1.00 and output $12.00. All three lines exactly double.

The part that costs money is how it applies. xAI's wording is that models with long context pricing bill the long context rates for all tokens in a request once its prompt reaches the threshold. The trigger is the prompt. The effect is the whole request, including every output token, and output is the line that moved to $12.00. This is not a surcharge on the tokens past 200,000, it is a different price for the entire call, and it arrives in one step.

Put a number on that step. Take a 199,999-token prompt with 5,000 output tokens and no caching: $0.399998 of input plus $0.03 of output is $0.429998. Add one token to the prompt and the same call bills $0.80 plus $0.06, or $0.86. That single token cost 43 cents. On Bedrock's Global card the same token costs $0.000002, because nothing happens there at 200,000. The ratio between the two is 215,001 to one, which is the sort of number that only exists when a price is a step function and somebody is standing on the step.

We wrote up the threshold itself when the model launched, in our August 15 post on Grok 4.6's rate card. What we did not know then, because three of these four venues did not exist yet, is that the threshold is the only thing anybody disagrees about.

Who carries the tier, and who does not

Input, cached input and output per million tokens, for a request whose prompt is 260,000 tokens long. Two venues double, two do not, and one of the two that does not cannot accept the request in the first place.

Azure has no long-context price because Azure has no long context

Microsoft's Foundry blog published the rate card on August 26 as a single row: Grok 4.6, Global Standard, $2.00 input, $6.00 output, $0.50 cache. No second column, no footnote, no threshold. Read on its own that looks like Microsoft declining to charge a surcharge that xAI charges, which would make Azure the cheapest way to run a 400,000-token prompt on this model.

It is not, because you cannot send one. Microsoft's own model documentation lists grok-4.6 with an output maximum of 128,000 tokens and a context window of 200,000 tokens. Not 500,000. xAI, AWS and OpenRouter all publish 500,000 for the same model and Google's model page publishes 524,288, so Microsoft is alone by a factor of two and a half, at exactly the number where the surcharge would have started.

So the flat card is honest and also uninformative. There is no tier on Azure's table for the same reason there is no tier on a table of Grok 4.6 prices between 600,000 and 700,000 tokens. The region does not exist. What Azure is selling is the cheap half of xAI's card with the expensive half fenced off, at 40% of the window for 100% of the headline price, and Microsoft does not connect those two facts anywhere we could find. Whether the cap is a preview limitation or a deliberate product decision, Microsoft has not said; the model has been in public preview for four days.

One more thing about Azure worth knowing before you plan around it. Grok 4.6 does not appear on Microsoft's official Azure pricing page at all. That page's xAI section stops at Grok 4.3. The only place Microsoft has published a number for this model is a blog post, which is a weaker document to be billed against than a price list.

Google copied the tier, then wrote the boundary down differently

Vertex is the venue that reproduces xAI's structure rather than flattening it. Google's pricing page carries the Grok models in two columns, one headed 200K input tokens or fewer and one headed more than 200K, and Grok 4.6 reads $2.00 against $4.00, $6.00 against $12.00, and $0.50 against $1.00. Same threshold, same doubling, same all-tokens rule.

The two do not quite agree on where the line is. Google's footnote under the Grok table says long context rates apply if a query input context is longer than 200K tokens. xAI labels its row long context at 200k tokens and above. At 199,999 tokens both charge the cheap card and at 200,001 both charge the expensive one, so the disagreement is one token wide, sitting exactly on 200,000. A prompt of precisely that length, with 8,000 output tokens and no caching, bills $0.4480 on Vertex and $0.8960 on xAI for the same work.

Nobody is going to hit that on purpose and we are not going to pretend it is a material saving. It is worth a sentence because of what it says about the documents: these are four independent transcriptions of one company's pricing decision, and they have already drifted at the edges. Google's own page proves the point against itself, since the Anthropic table a few rows above the Grok one uses the phrase longer than or equal to 200K tokens for what is otherwise the identical rule.

What a long-context agent turn actually bills

Here is one request priced on every card. A coding agent sends 260,000 tokens of repository context, 85% of it served from cache because the same files come back every turn, and generates 8,000 tokens of output. That is an ordinary turn for this class of tool and it sits 60,000 tokens over the line.

The spread across one model on one request is four times, from $118.25 to $473.00 per thousand turns, before you count the tier that will not take the call. Nothing about the model changes across those rows. Same weights, a window of 500,000 tokens or more on every row that can serve the request, same reasoning efforts.

Shorten the prompt to 199,999 tokens and the picture collapses. xAI bills $0.1930 for that turn and Bedrock Global bills the same $0.1930, because below the line their cards are character for character identical. The entire argument of this post lives in the 300,000-token band above the threshold, and if your prompts stay under 200,000 you can stop reading and buy from whoever you already have a contract with.

AWS charges a premium that is cheaper than xAI charging nothing

Bedrock sells three service tiers on this model. Standard is the card above. Flex is half of it, for work you are willing to have queued. Priority is 1.75x, for work you are not. On the Global card Priority comes to $3.50 input, $0.875 cached and $10.50 output.

Line those up against xAI's long-context card of $4.00, $1.00 and $12.00 and Priority is 0.875x on all three, which is to say AWS's most expensive way to buy Grok 4.6 undercuts xAI's ordinary way to buy it by 12.5% once the prompt is long. Flex on the same card is 0.25x. That is a quarter of the price, from the reseller, for the model xAI trained.

xAI has a premium tier of its own and it does not help. Priority Processing is a flat 2x on every token type, applied after the cache discount, and it stacks on top of the long-context rates rather than replacing them, so a long prompt in a hurry bills $8.00 and $24.00 per million. There is also no batch discount to fall back on. xAI publishes a 20% batch rate for Grok 4.3 and the 4.20 models and then states that models not listed have none, which excludes 4.6 by leaving it out. The model JSON behind that page carries a batch discount of 0 for it, so the omission is deliberate rather than an oversight.

Caching moves you toward the cliff, not away from it

The intuitive defence against a prompt-length threshold is to cache the prefix, on the theory that cached tokens are cheap tokens and cheap tokens should not count for as much. xAI closes that door explicitly. Its caching documentation states that long context pricing applies when total prompt tokens including cached tokens exceed the threshold, and that both cached and non-cached tokens then use their respective long-context rates.

Which inverts the usual advice. On this model a big cached prefix is the thing most likely to put you over 200,000 tokens, and the moment it does, the cache reads you were relying on go from $0.50 to $1.00 per million along with everything else. Caching still saves money, it just saves it inside the expensive tier rather than keeping you out of one. An agent that grows its context every turn is walking toward the line by design, and the cache is what carries it there.

That is the mechanism behind the 85% cache hit rate in the table above producing a $473.00 bill rather than a cheap one. It is also why this threshold behaves differently from the GPT-5.6 long-context tier we looked at on August 9, and from Gemini 3.1 Pro Preview, whose own 200K step raises output 1.5x rather than doubling it.

One number we think is wrong, and it is not AWS's

OpenRouter lists five endpoints for this model, four from xAI and one from Amazon Bedrock in us-west-2. Every one of the five carries an override keyed to a minimum prompt length of 200,000 tokens that doubles all three rates. Including the Bedrock endpoint, which OpenRouter prices at $2.20, $0.55 and $6.60 rising to $4.40, $13.20 and $1.10.

All three of those base figures match AWS's In-Region card exactly, so OpenRouter has read the right price list. The doubling does not appear on it. AWS's model card is a three-row table with a single note that all prices are per million tokens for the Standard tier, and it publishes a 500,000-token context window beside it with no qualification. Our reading is that OpenRouter applies the tier at the model level, where it is correct for the four xAI endpoints, and inherits it onto the Bedrock row where it is not. We could be wrong in the other direction and AWS could be documenting its price incompletely, which would be worse. Either way, one of two companies is publishing a number you would be billed against and it does not match the other.

Worth noting alongside it: the Bedrock endpoint is the only one of the five with implicit prompt caching turned on, and its cache write price is $0. The four xAI endpoints report implicit caching false.

The launch post quotes $2 and $6 and stops

xAI's announcement for this model says pricing starts at $2 per million input tokens and $6 per million output tokens, and adds that there is a fast variant at twice the price. The words starts at are carrying the entire long-context tier, which the page never mentions. The fast variant is almost certainly Priority Processing, the only 2x option xAI documents, though neither page says so; OpenRouter's two priority endpoints price at exactly twice the standard card, which is the closest thing to confirmation on offer.

We are not calling that dishonest. Starts at is accurate and $2.00 genuinely is the entry price. But it is the sentence that every downstream tracker copied, and it is why a model with a 500,000-token window and a doubling at 40% of it is filed everywhere as a $2 model. The benchmarks that page publishes are the ones people quote alongside it:

xAI's own reported figures at high reasoning effort, from the launch post. On the same page, GPT-5.6 Sol Max scores 34.6% and Fable 5 Max 34.1% on Terminal-Bench v3.0 against Grok 4.6's 26%, so the model xAI is comparing itself to beats it on the hardest agentic row it chose to publish. Artificial Analysis puts the Intelligence Index at 61 and the cost of running its full evaluation suite at $1,157.64 across 72M output tokens.

One number decides this, and most teams do not know theirs

If your prompts live under 200,000 tokens, buy from anyone. The cards are identical and the only lever left is Bedrock Flex at half price if latency does not matter to you.

If they cross it, the venue is worth twice the model. Bedrock Global is half of xAI and Vertex on the same request, Bedrock In-Region is 55% of them if you need the data to stay in one region, and Azure is not an option at all no matter what its flat rate card implies. Measure where your prompts actually sit before you decide, because the entire decision is a single number and most teams do not know theirs.

And if you are already on xAI direct with long prompts, the cheapest change available to you is not a smaller model or a better cache. It is moving the same calls to a reseller, which is a sentence we have not been able to write about any other frontier model this year.

Microsoft published its price on a blog. That is the weakest document on this page.

• xAI: pricing - Footer reads last updated August 21, 2026. The grok-4.6 row gives 500k context, $2.00 / $0.50 / $6.00 short and $4.00 / $1.00 / $12.00 long, labelled long context at 200k tokens and above. Source of the all-tokens rule, of Priority Processing at a flat 2x applied after the cache discount, and of the batch table that lists 20% for Grok 4.3 and the 4.20 models and then says models not listed have no batch discount
• xAI: prompt caching usage and pricing - The sentence the caching section of this post rests on, stating that long context pricing applies when total prompt tokens including cached tokens exceed the threshold and that both cached and non-cached tokens then use their long-context rates
• AWS: Bedrock model card for Grok 4.6 - Three rows, In-Region and Geo CRIS at $2.20 / $6.60 / $0.55 and Global CRIS at $2.00 / $6.00 / $0.50, per million tokens, Standard tier, with no context tier anywhere on the page. Also the source of Priority at 1.75x and Flex at 0.5x, the 500K window, implicit prompt caching, and the four reasoning efforts. The launch date is AWS's own, given as August 18, 2026; xAI's companion post at x.ai/news/grok-4-6-amazon-bedrock is dated August 19 and calls it generally available, so the two vendors disagree by a day and we have used AWS's
• Google: Vertex AI generative AI pricing - The xAI Grok section, two columns headed 200K input tokens or fewer and more than 200K, with Grok 4.6 at $2.00 / $6.00 / $0.50 against $4.00 / $12.00 / $1.00. The Grok footnote reads longer than 200K while the Anthropic footnote higher on the same page reads longer than or equal to 200K, which is where the one-token disagreement comes from. Read from the raw page on August 30, 2026 because the table renders below the point our fetcher truncates. Google's own model card at docs.cloud.google.com gives the release date as August 21, 2026 and the context length as 524,288 rather than the round 500,000 the other three publish
• Microsoft: Grok 4.6 comes to Foundry Models - August 26, 2026, and the only place Microsoft has published a price for this model. One row, Global Standard, $2.00 / $6.00 / $0.50, no tier. The official Azure pricing page for xAI models stops at Grok 4.3 and has no Grok 4.6 entry at all
• Microsoft: Foundry Models sold directly by Azure - The document this post turns on. The grok-4.6 row reads output text 128,000 tokens max and context window 200,000 tokens, against the 500,000 every other venue publishes. Grok 4.6 appears only under Global Standard, in no DataZone, Regional or Provisioned table
• OpenRouter: Grok 4.6 endpoint detail - Pulled August 30, 2026. Five endpoints, four xAI and one Bedrock us-west-2, every one carrying a 200,000-token minimum prompt override that doubles all rates, and every one with a discount field of 0. The Bedrock row's doubling is the contradiction described above. Also the source of the 450,000-token max completion figure, the reasoning defaults, and the SpaceXAI branding on the model name
• xAI: Grok 4.6 announcement - The pricing starts at $2 sentence, the unnamed fast variant at twice the price, and the benchmark table quoted above. Companion posts at x.ai/news/grok-4-6-amazon-bedrock and x.ai/news/grok-4-6-vertex-ai carry the two cloud launches
• TokenCost: Grok 4.6 pricing - Our August 15 post on xAI's own card, which covers the 67% cache increase over Grok 4.5 and the threshold mechanics in more depth than this post repeats. Our catalogue entry has been updated alongside this one to carry the four venues
• What we could not establish. Whether Azure's 200,000-token window is a preview limitation or permanent, and whether it is the cause of the flat card or a coincidence, are our inference from two Microsoft documents that never reference each other. We could not resolve the OpenRouter and AWS contradiction from primary sources and have described both rather than picking one. Microsoft's blog reproduces three benchmarks as images footnoted as coming from the model provider, so we did not transcribe them. The Artificial Analysis figures came through a rendered page rather than raw markup and are the softest numbers here. And every price in this post is a list price read on August 30, 2026: none of these venues publishes committed-use or enterprise discounting, which is exactly the layer where a four-times spread on paper usually stops being a four-times spread in practice