How much does Grok 4.6 cost on Amazon Bedrock?

Bedrock in-region and geo cross-region requests cost $2.20 per million input tokens and $6.60 per million output tokens, with cached reads at $0.55. Global cross-region inference matches xAI's direct rate exactly at $2.00 input and $6.00 output per million tokens, with cached reads at $0.50. Grok 4.6 went generally available on Bedrock on August 19, 2026.

Is Grok 4.6 cheaper direct from xAI than on Bedrock?

Direct from xAI costs $2/$6 per million tokens. Bedrock's global cross-region inference matches that exactly, so there is no premium if you can route globally. Bedrock in-region and geo cross-region carry a 10% premium at $2.20/$6.60, which is the standard cost of pinning inference to a region for data residency.

What is the Grok 4.6 context window on Bedrock?

500,000 tokens. Grok 4.6 on Bedrock accepts text and image input, returns text output only, and supports four configurable reasoning effort levels — low, medium, high and xhigh. It is available in every AWS Region where Bedrock is offered.

Quick Answer

Grok 4.6 on Bedrock vs xAI Direct API: Cost Compared

Published: August 20, 2026

The Short Answer

If you can use global cross-region inference, Bedrock costs exactly the same as going direct to xAI. If you need regional pinning, you pay 10%.

Route	Input /MTok	Output /MTok	Cached read
xAI direct API	$2.00	$6.00	$0.50
Bedrock global cross-region	$2.00	$6.00	$0.50
Bedrock in-region / geo cross-region	$2.20	$6.60	$0.55

Grok 4.6 became generally available on Amazon Bedrock on August 19, 2026. Verified August 20, 2026.

What Landed On August 19

xAI’s Grok 4.6 — released direct on August 12, 2026 — reached Amazon Bedrock a week later with the full feature set:

500,000-token context window
Four reasoning effort levels: low, medium, high, xhigh
Text and image input, text-only output
Available in every AWS Region where Bedrock is offered
Support for long-running agents and interactive work

The reasoning-effort control is the underrated part. Most cost blowups on reasoning models come from paying frontier-grade thinking tokens on tasks that never needed them. Four discrete levels means you can tier your workload rather than paying xhigh rates on classification.

The Pricing Structure, Properly Explained

AWS’s three-tier routing is the whole story, and it confuses people constantly.

Global cross-region inference lets AWS route your request to whichever region has capacity, anywhere. It is the cheapest tier — $2.00/$6.00, identical to xAI direct — because AWS gets scheduling flexibility in exchange.

Geo cross-region inference constrains routing to a geography (for example, EU regions only). $2.20/$6.60.

In-region pins inference to one specific region. Also $2.20/$6.60.

So the real question is not “Bedrock or xAI.” It is “do I need to control where inference physically happens?” If the answer is no, Bedrock is free of premium. If the answer is yes — GDPR data residency, a regulated sector, a contractual commitment to customers — the 10% is what that control costs, and it is cheap relative to the alternative of not being able to use the model at all.

Where Grok 4.6 Sits On Price

Against the frontier tier as of August 2026 (per MTok, input/output):

Model	Input	Output	30K-in/5K-out task
Grok 4.6	$2	$6	$0.09
Claude Opus 5	$5	$25	$0.275
GPT-5.6 Sol	$5	$30	$0.30
GPT-5.6 Terra	$2	$12	$0.12
Gemini 3.7 Flash	$0.75	$3.75	$0.041
GLM-5.3	$1.40	$4.40	$0.064

Grok 4.6 is roughly a third the cost of Opus 5 or Sol on a standard task shape while being positioned in the same frontier conversation. That is the reason it keeps appearing in cost-driven stacks.

The Caveat Nobody Mentions

Grok 4.6’s headline rate applies to prompts under 200K tokens. Crossing 200K reprices the whole request, not just the overage. With a 500K window available on Bedrock, this is an easy and expensive mistake to make: a long-context agent run that drifts past the threshold does not cost 10% more, it reprices entirely.

If you are building agents that accumulate context — and most agent harnesses do — either enforce a hard context budget below 200K, or model your costs at the higher tier from the start. Cached input on Grok 4.6 is $0.50 (up from $0.30 on Grok 4.5), so aggressive prompt caching helps but does not eliminate the cliff.

When To Choose Bedrock

Choose Bedrock if you already have AWS commitments and want Grok inside existing IAM, VPC, CloudWatch and billing; if you need data residency guarantees; or if you want one API surface across Claude, Nova and Grok so switching models is a config change. The consolidation argument is stronger in August 2026 than it was six months ago — after OpenAI paused frontier training on August 18 and Anthropic restricted Claude Fable 5 access earlier in the year, capability-triggered availability risk is real, and a single abstraction over multiple providers is genuine insurance.

Choose xAI direct if you are not on AWS, you want the newest features on day one rather than after Bedrock integration, or you want the simplest possible billing relationship. Direct access also got Grok 4.6 seven days earlier — that gap is the recurring cost of the managed route.

Last verified: August 20, 2026.