Grok 4.6 landed on Amazon Bedrock — and AWS just published a price for data residency: 10%

What actually shipped

AWS announced Grok 4.6 on Amazon Bedrock on 19 August 2026; the model card lists a launch date of 18 August. In AWS’s description it is a model “built for coding, agentic tasks, and knowledge work” — wording lifted almost verbatim from SpaceXAI, whose own post calls it “our latest flagship model, built for long-running agents and ambitious interactive and visual work.”

The specifications match what shipped on 12 August: a 500K-token context window, configurable reasoning effort across low, medium, high and xhigh, text and image input, text output. Reasoning is always active, with low as the default. Access is via the Responses, Chat Completions and Converse APIs.

Regional coverage is broad. Global cross-Region inference reaches roughly thirty Regions across the US, Canada, Europe, Asia-Pacific, the Middle East, Africa and South America. Geo cross-Region inference — US-scoped — covers four US Regions. That is a normal, unremarkable Bedrock launch, and if the story stopped there it would not be worth your time.

The pricing table is the story

Here is what the model card publishes, per million tokens, Standard tier:

Now set that against xAI’s own published rate for prompts under 200K tokens: $2.00 input, $6.00 output, $0.50 cached.

The Global rate is not close to xAI’s list price. It is xAI’s list price, to the cent, cache reads included. Which means the usual assumption about model marketplaces — that the cloud takes a cut and you pay a convenience premium — is simply false here. Buy Grok 4.6 through Bedrock with global routing and you pay exactly what you would pay xAI directly, while inheriting Bedrock’s invocation logging, CloudWatch metrics and Cost Explorer attribution for nothing.

The premium appears only when you constrain where the request may go. Geo cross-Region inference keeps requests inside US geography for data-residency purposes. It costs 10% more. In-Region — a single Region, the strictest option — costs the same 10% more.

That is worth sitting with, because of what it makes visible. Data residency is normally invisible in pricing. It shows up as a Region you select, a clause in an enterprise agreement, a line in a security questionnaire. Here it is a published per-token rate. AWS is telling you, in a public docs table, what it costs to promise that your tokens stay in a geography: one tenth of the bill.

The mechanism is not mysterious. Global routing lets Bedrock serve a request from wherever it has capacity; a residency constraint shrinks that pool and makes capacity management harder. The surcharge is the cost of the constraint, and 10% is arguably cheap for it. But it converts a compliance requirement into a number, and numbers get compared. A team told to keep inference in the US can now say precisely what that instruction costs — and a team that assumed it needed residency may look harder at whether it actually does.

One further note for anyone modelling spend: Grok 4.6 on Bedrock is offered on the Standard service tier only. The model card marks Priority, Flex and Reserved as unavailable. There is no committed-throughput discount and no batch tier to grow into, so the rate above is the rate — a meaningful difference from how enterprises usually plan model spend as volume scales.

”Grok 4.6 on Bedrock” is two different products

The second finding is buried in the feature matrix, and it will bite someone.

Bedrock exposes this model through two endpoints, and they do not support the same things.

On bedrock-runtime — the endpoint most teams will reach for, and the one with broad regional coverage — the model card marks as supported: invocation logs, response streaming, prompt caching, reasoning, and projects (default project only). It marks as not supported: server-side tool use, intelligent prompt routing, count tokens, structured outputs, and application inference profiles.

On bedrock-mantle, the picture reverses. Client-side tool calling, reasoning, projects, abuse detection, response streaming, structured outputs and prompt caching are all supported, with nothing listed as unsupported.

So if your application needs structured outputs, bedrock-mantle is your endpoint. Except that Grok 4.6 is available on bedrock-mantle in exactly one Region — us-west-2, Oregon — as In-Region inference only, with Geo and Global both unavailable. Structured outputs with Grok 4.6 on Bedrock therefore means Oregon, no cross-Region failover, and no data-residency story outside the United States.

There is a matching asymmetry on the runtime side worth knowing before you write code: on bedrock-runtime, Grok 4.6 is not available for In-Region inference at all. You must name a cross-Region inference profile — us.xai.grok-4.6 or global.xai.grok-4.6 — as the model ID rather than the bare xai.grok-4.6. And your IAM identity needs bedrock:InvokeModel on your account’s default project in addition to the inference profile, which is exactly the kind of detail that turns a fifteen-minute integration into an afternoon.

Reasoning has its own wrinkle. Reasoning content is returned encrypted, retrievable by passing include: ["reasoning.encrypted_content"] on the Responses API and passable back in later turns for multi-turn context. The Chat Completions API does not return reasoning tokens at all. If you were planning to inspect or log the model’s reasoning for evaluation or audit purposes, the API you choose determines whether you can.

The 200K cliff, and the tool that won’t count for you

The sharpest open question is about long context, and it deserves care because the honest answer is unresolved.

xAI’s documentation prices Grok 4.6 in two bands. Under 200K prompt tokens: $2.00 / $0.50 / $6.00 for input, cached input and output. At or above 200K: $4.00 / $1.00 / $12.00. And then the sentence that matters, in xAI’s own words — requests whose prompt reaches the listed token threshold are “billed at the higher rate for all tokens in the request.”

Not the excess. All of them. A 201K-token prompt costs roughly double a 199K-token one, and the step is invisible until the invoice arrives.

This matters far more than it would have a year ago, because the workload Grok 4.6 is explicitly sold for — long-running agents, large repositories, extended sessions — is precisely the workload that accumulates context until it crosses a line nobody is watching. A 500K context window advertised alongside a 200K billing threshold is a genuine trap, and it exists on xAI’s direct API regardless of what Bedrock does.

What Bedrock does is the unresolved part. The model card publishes flat Standard-tier rates with no long-context band. That could mean AWS absorbed the tiering into a single blended rate; it could mean the band exists and is documented elsewhere; it could mean something else. The card does not say, and inferring a favourable answer from an absence is exactly the reasoning that produces surprise invoices. Confirm it with AWS before you budget a long-context workload, and treat any assumption either way as unverified.

The reason this is more than a footnote: bedrock-runtime does not support the count-tokens API. On the endpoint with broad regional availability, you cannot ask Bedrock how many tokens your prompt contains before you send it. If a threshold does apply, the platform will not help you see the edge you are walking toward. Anyone running agents at scale should be counting tokens client-side and capping context deliberately — which is good practice anyway, and here it is close to mandatory.

Why this matters

Procurement objections just disappeared, and that changes the shortlist. The strongest argument against Grok in an enterprise was never capability — it was that adopting it meant onboarding a new vendor, negotiating a new contract, and explaining a new data path to a security team. Bedrock erases all three. Grok 4.6 now arrives under an AWS agreement most enterprises already hold, and this is the same mechanism that mattered when OpenAI’s models reached Bedrock GA in June. The model did not get better. Its addressable market did.

Data residency now has a comparable number attached to it. A published 10% surcharge invites a question that used to be unaskable: is our residency requirement a legal obligation or an inherited habit? For genuinely regulated workloads the answer is settled and 10% is a bargain. For the substantial category of teams who specify US-only routing because it seemed prudent, a line item makes the cost of prudence explicit — and once one provider prices the constraint, others get asked to. Update, 23 August 2026: that took three days. Claude Code v2.1.239, released 22 August, now folds a 1.1× US-only-inference premium for data-residency workspaces into its own cost estimates — /cost, the status line and --max-budget-usd all reflect it. Two vendors, two independent pricing decisions, the same number: residency costs 10%. That is no longer one marketplace’s quirk; it is an emerging market rate, and it is now surfacing inside the developer tooling rather than only on a pricing page. If you run Claude Code in a residency workspace, your displayed budget just got more honest — and any cost model built before 22 August is now understating spend by a tenth.

Marketplace listings are not equivalent to direct API access, and the gaps are load-bearing. Structured outputs, tool-use modes and token counting are not exotic extras; they are how production applications are built. A team that benchmarks Grok 4.6 against Claude or ChatGPT on quality alone, picks it on price, and then discovers the runtime endpoint won’t return structured output has made an expensive mistake with a cheap fix — read the feature matrix per endpoint before the bake-off, not after.

The distribution picture around SpaceX keeps getting stranger. SpaceXAI’s flagship is now sold through Amazon’s marketplace, while SpaceX-owned Cursor ships a Git forge competing with GitHub and Warp bets on model-agnostic agent fleets. Vertical integration at the application layer, horizontal distribution at the model layer. Both bets are being placed by the same corporate parent, in the same week.

The verdict

Grok 4.6 reaching Bedrock is unremarkable as news and useful as a document. The model is what it was on 12 August: a strong-value near-frontier option that leads on knowledge work and trails on agentic coding, worth choosing on cost-per-task rather than on capability. Nothing about the AWS listing changes that assessment, and the broader price competition is where the real pressure on these numbers comes from.

What the listing does is publish three things the industry usually keeps implicit: that a cloud marketplace can resell a frontier model at the lab’s exact list price, that a residency guarantee has a per-token cost, and that “available on Bedrock” describes a feature surface that varies by endpoint and Region.

Recommendation: If you already run on AWS and want Grok 4.6, use Global cross-Region inference — you pay xAI’s own list price and get Bedrock’s logging and cost attribution free. Choose the Geo profile only if a residency requirement genuinely binds you, and now that the surcharge is 10% and visible, make someone confirm that it does. Before any production commitment, do two checks: verify the long-context billing behaviour with AWS in writing, because a 500K window sitting above a 200K doubling threshold is a real budget risk on an endpoint that will not count tokens for you; and confirm your required features against the specific endpoint you intend to call, because structured outputs pin you to a single Region. If either check comes back wrong for your workload, going direct to xAI remains the cleaner path — and if you are choosing between coding assistants rather than raw APIs, the Cursor versus Claude Code question and our best AI coding tools roundup are the more relevant comparisons.

Update, 21 August 2026: the “who sees your tokens, and what do they charge for it” question moved up a layer the following day. Stripe agreed to acquire OpenRouter at a reported $7.5 billion and Ramp launched a free rival gateway whose default is to retain inputs, outputs and tool calls for a year. If you are pricing a data-residency position on Bedrock, price the gateway hop too — the neutral routing layer just changed owners.

Questions or corrections? Email Pick Right. Want the full list? See all news.