xAI’s Grok 4.6 is now obtainable in Amazon Bedrock

In the present day, we’re saying that xAI’s Grok 4.6 is accessible in Amazon Bedrock, including a frontier mannequin constructed for long-running brokers, coding, and information work to the Bedrock mannequin catalog. Grok 4.6 launched on Bedrock on August 18, 2026. It gives a 500K token context window and helps configurable reasoning effort at 4 ranges: low, medium, excessive, and xhigh.

That is xAI’s second mannequin in Amazon Bedrock. When Grok 4.3 turned usually obtainable, xAI joined Amazon Bedrock as a mannequin supplier and the mannequin was reachable by Bedrock Mantle, the OpenAI-compatible inference engine in Amazon Bedrock. Grok 4.6 widens that floor space significantly: it’s obtainable on each the bedrock-mantle and bedrock-runtime endpoints, and it helps the Converse API alongside Chat Completions and Responses.

This publish covers what xAI says Grok 4.6 is designed for, how it’s packaged on Amazon Bedrock, and methods to ship your first request.

What Grok 4.6 is constructed for

The aptitude and coaching particulars on this part come from xAI’s launch announcement, Introducing Grok 4.6.

Grok 4.6 builds on Grok 4.5 with a selected give attention to long-running brokers and extra formidable interactive and visible work. xAI describes the mannequin as staying with complicated duties throughout many steps, whether or not that’s researching a subject, analyzing info, working throughout a code base, or turning an concept into a elegant software or work artifact.

On coaching, xAI stories an extended supplemental coaching run than Grok 4.5, utilizing curated model-generated information for reasoning and superior technical ideas, high-quality engineering information, and an improved optimizer and coaching recipe. It then used Grok 4.5 to regenerate the supervised fine-tuning trajectories throughout reasoning efforts, agent harnesses, and domains together with STEM, software program engineering, and information work, filtering out problematic traces with model-based checks. The mannequin was then educated on a variety of agentic reinforcement studying duties spanning information work, common coding, and domain-specific environments reminiscent of kernel optimization, internet growth, and computer-aided design.

Two behaviors xAI calls out are price noting for anybody constructing brokers. On longer trajectories, the mannequin started exhibiting extra self-testing and verification, checking its personal work earlier than shifting on. It additionally produces stronger first passes on visible and interactive initiatives, establishing the construction and visible language of an software in a single move, which the crew discovered helpful the place the quickest path to a very good consequence was to start out with one thing substantial after which iterate.

On security, xAI states that Grok 4.6’s safeguards have been improved and calibrated in keeping with the mannequin’s capabilities, backed by what it describes as its widest-ever suite of pre-deployment testing for capabilities and safeguard calibration, plus post-deployment and third-party testing. The corporate positions its security stack as maximizing utility and safety throughout reputable use instances in domains reminiscent of vulnerability patching, accelerating the engineering design cycle, and augmenting AI analysis.

Reported benchmark outcomes

xAI stories that Grok 4.6 achieves frontier intelligence throughout a number of agentic coding and information work benchmarks. These are the figures it revealed for Grok 4.6 Excessive at launch on August 12, 2026:

Supply: xAI, in line with https://x.ai/news/grok-4-6.

A number of of these evaluations come from Synthetic Evaluation, so it helps to know what they measure. Based on Artificial Analysis, the Synthetic Evaluation Intelligence Index v4.1.1 is a composite that comes with 9 evaluations: GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity’s Final Examination, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR. These cowl agentic device use, reasoning and information, information reliability, lengthy context reasoning, and quantitative evaluation over spreadsheets and paperwork. AA-Briefcase is its agentic information work benchmark, the place AA-Briefcase Elo aggregates rubric move charge, analytical high quality Elo, and presentation Elo, with larger scores higher.

Synthetic Evaluation additionally tracks price and latency alongside intelligence. Its cost-per-task metric is a weighted common price per Intelligence Index process, derived from enter, cache hit, cache write, reasoning, and reply token costs, which is a helpful lens in case you are sizing a reasoning-heavy agent workload the place reasoning tokens are an actual line merchandise.

What Grok 4.6 provides on Bedrock

A number of Bedrock capabilities are new for this mannequin relatively than carried over from the sooner Grok launch.

The bedrock-runtime endpoint. Grok 4.6 is served on bedrock-runtime along with bedrock-mantle, so you possibly can attain it with the AWS SDKs and the usual Bedrock management floor relatively than solely an OpenAI-compatible shopper.

The Converse API, together with streaming. Each converse and converse_stream can be found. That is the sensible payoff of runtime assist: one message form throughout fashions, and streaming by the same old Converse occasions (messageStart, contentBlockDelta, contentBlockStop, messageStop, metadata) with out hand-rolling server-sent occasions (SSE) parsing.

An xhigh reasoning effort stage. Effort runs low, medium, excessive, xhigh, extending the vary on the prime finish for issues the place a deeper move is definitely worth the tokens. On Converse, set it by additionalModelRequestFields={"reasoning_effort": "xhigh"} relatively than a reasoning parameter.

Cross-Area inference. On bedrock-runtime you route by considered one of two inference profiles relatively than pinning to a single Area. us.xai.grok-4.6 retains site visitors throughout the US geography when you’ve got information residency necessities, and world.xai.grok-4.6 routes worldwide for the widest capability pool. International can be the cheaper of the 2, at $2.00 per million enter tokens in opposition to $2.20, so absent a residency constraint it’s normally the higher default.

Amazon Bedrock Guardrails. Grok 4.6 now helps Guardrails on bedrock-runtime throughout its APIs, providing you with content material filters, denied subjects, personally identifiable info (PII) redaction, and phrase insurance policies. You connect a guardrail by ID and model on the request, and the coverage is evaluated in opposition to each the immediate and the mannequin’s response. For agentic workloads this issues as a result of it places a constant coverage boundary round a mannequin which may run unattended throughout many steps.

Invocation logging. With mannequin invocation logging enabled, Grok 4.6 calls are captured as full Amazon CloudWatch information: request physique, response physique, token counts together with reasoning tokens, and the inference profile used. Helpful for auditing agent runs the place you could see what the mannequin was truly requested.

Immediate caching. Cached enter is billed at roughly 1 / 4 of the usual enter charge, which issues for brokers that resend a big system immediate or doc on each flip. Caching applies to a repeated prefix, so maintain secure content material on the entrance of the request, and skim the cached token depend within the utilization block to verify the low cost is touchdown earlier than you construct it into a price mannequin.

Device calling, structured output, picture enter, response streaming, and encrypted reasoning content material can be found as properly, however these date from the Grok 4.3 launch and are lined in that publish.

How Grok 4.6 is packaged on Amazon Bedrock

Grok 4.6 accepts textual content and picture enter and returns textual content. Audio, speech, video, and embedding modalities aren’t supported, and it doesn’t generate pictures. The mannequin is reachable by two endpoints, and the mannequin ID differs relying on which one you utilize:

On the API aspect, Grok 4.6 helps the Responses API, the Chat Completions API, and the Converse API. The Invoke API isn’t supported.

Characteristic assist differs by endpoint, which is the element most certainly to form your integration selection:

On bedrock-mantle, supported options embody client-side device calling, reasoning, structured outputs, immediate caching, response streaming, initiatives, and abuse detection.

On bedrock-runtime, supported options embody reasoning, immediate caching, response streaming, invocation logs, and initiatives (default challenge solely). Structured outputs, server-side device use, clever immediate routing, depend tokens, and software inference profiles aren’t supported on that endpoint.

Device calling works on each endpoints. The mannequin returns a structured operate request, your code executes it, and also you move the consequence again. On bedrock-runtime you possibly can drive that loop by Converse’s toolConfig or the OpenAI-compatible instruments parameter, so brokers that depend upon operate calls aren’t restricted to bedrock-mantle.

In case your software will depend on JSON Schema structured output, that factors you at bedrock-mantle. If you need the Converse API or invocation logging, that factors you at bedrock-runtime.

Areas and inference choices

Availability differs by endpoint. On bedrock-mantle, Grok 4.6 is accessible for in-Area inference in US West (Oregon) (us-west-2) . On bedrock-runtime, in-Area inference isn’t supplied. As an alternative, you invoke the mannequin by cross-Area inference profiles. Geo cross-Area inference is accessible from the US Areas (us-east-1, us-east-2, us-west-1, and us-west-2), and International cross-Area inference is accessible from a significantly longer record spanning the US, Canada, Europe, Asia Pacific, the Center East, Africa, and South America. Geo cross-Area routes throughout Areas inside a geography whereas respecting information residency, and International cross-Area routes wherever worldwide when there are not any residency constraints. The total desk runs to greater than 30 Areas, so test the mannequin card and the Regional availability by mannequin web page for the present record earlier than you pin a Area.

This can be a change in form from the Grok 4.3 launch, the place, as famous within the Grok 4.3 publish, the mannequin used in-Area inference solely and Geo and International cross-Area inference weren’t supplied.

Service tier and pricing

Grok 4.6 helps three service tiers. Customary is pay-per-token with no dedication, chosen by setting "service_tier": "default" or omitting the sector. Precedence delivers sooner, prioritized processing for a premium ("service_tier": "precedence"). Flex gives lower-cost entry for work that isn’t time-sensitive ("service_tier": "flex"). For per-token pricing throughout the tiers, see the Amazon Bedrock pricing web page.

The opposite two tiers are priced as multipliers on these Customary charges: Precedence at 1.75x, a 75 p.c premium, and Flex at 0.5x, a 50 p.c low cost. So the identical workload that prices $2.20 per million enter tokens on Customary in-Area runs $3.85 on Precedence and $1.10 on Flex, which makes tier choice a bigger price lever than the Area selection.

For reference, xAI lists Grok 4.6 pricing beginning at $2 per million enter tokens and $6 per million output tokens, with a quick variant at twice the value. At all times affirm present charges on the Amazon Bedrock pricing web page, as a result of costs and tiers change.

Ship your first request

Earlier than your first name, affirm the mannequin is accessible to you within the Bedrock console for the Area you intend to make use of. Grok 4.6 is served by inference profiles relatively than on-demand throughput on the naked mannequin ID, which is why requests title us.xai.grok-4.6 or world.xai.grok-4.6 on bedrock-runtime.

Grok 4.6 makes use of OpenAI-compatible APIs, so the OpenAI SDK works in opposition to both endpoint after you set the bottom URL. Set up the SDK, and boto3 in case you plan to make use of the Converse API:

pip set up openai
pip set up boto3

Generate a long-term Amazon Bedrock API key from the Amazon Bedrock console for exploration, then set your setting. For bedrock-mantle:

export OPENAI_API_KEY="<present your Bedrock API key>"
export OPENAI_BASE_URL="https://bedrock-mantle.us-west-2.api.aws/openai/v1"

For bedrock-runtime:

export OPENAI_API_KEY="<present your Bedrock API key>"
export OPENAI_BASE_URL="https://bedrock-runtime.us-east-1.amazonaws.com/openai/v1"

A primary request on bedrock-mantle with the Chat Completions API:

from openai import OpenAI

shopper = OpenAI()

response = shopper.chat.completions.create(
    mannequin="xai.grok-4.6",
    messages=[
        {"role": "user", "content": "Can you explain the features of Amazon Bedrock?"}
    ],
)
print(response)

On bedrock-runtime the distinction is the mannequin title: you move a cross-Area inference profile as a substitute of the naked mannequin ID. This instance additionally switches to the Responses API to point out that form:

from openai import OpenAI

shopper = OpenAI()

response = shopper.responses.create(
    mannequin="us.xai.grok-4.6",
    enter="Are you able to clarify the options of Amazon Bedrock?",
)
print(response)

And thru the Converse API with boto3. As a result of reasoning is energetic, the primary content material block carries the reasoning and the reply sits in a later block, so search the blocks for the textual content relatively than indexing content material[0]:

import boto3

shopper = boto3.shopper("bedrock-runtime", region_name="us-east-1")

response = shopper.converse(
    modelId="us.xai.grok-4.6",
    messages=[
        {"role": "user", "content": [{"text": "Can you explain the features of Amazon Bedrock?"}]}
    ],
    inferenceConfig={"maxTokens": 2048},
)

blocks = response["output"]["message"]["content"]
textual content = subsequent(b["text"] for b in blocks if "textual content" in b)
print(textual content)

On Converse you set the hassle stage by additionalModelRequestFields relatively than a reasoning parameter:

response = shopper.converse(
    modelId="us.xai.grok-4.6",
    messages=[{"role": "user", "content": [{"text": "What is 17*23? Number only."}]}],
    inferenceConfig={"maxTokens": 3000},
    additionalModelRequestFields={"reasoning_effort": "xhigh"},
)

Three operational notes. First, on bedrock-runtime, Grok 4.6 isn’t obtainable for in-Area inference, so requests should title us.xai.grok-4.6 or world.xai.grok-4.6.

Second, bedrock:InvokeModel is evaluated in opposition to three sources: your account’s default challenge, the inference profile you title, and the underlying basis mannequin. The inspiration mannequin ARN is wildcarded throughout Areas as a result of cross-Area profiles route outdoors the calling Area. Bearer-token authentication on the OpenAI-compatible endpoints moreover requires bedrock:CallWithBearerToken, which boto3 and Converse don’t want:

{
    "Model": "2012-10-17",
    "Assertion": [
        {
            "Effect": "Allow",
            "Action": "bedrock:InvokeModel",
            "Resource": [
                "arn:aws:bedrock:{region}:{account-id}:project/default",
                "arn:aws:bedrock:{region}:{account-id}:inference-profile/us.xai.grok-4.6",
                "arn:aws:bedrock:*::foundation-model/xai.grok-4.6"
            ]
        },
        {
            "Impact": "Enable",
            "Motion": "bedrock:CallWithBearerToken",
            "Useful resource": "*"
        }
    ]
}

Checklist each inference profile you intend to name. Profiles are scoped individually, so a coverage naming us.xai.grok-4.6 doesn’t cowl world.xai.grok-4.6.

Third, the 2 authentication mechanisms cowl completely different code paths. An Amazon Bedrock API key in OPENAI_API_KEY travels as a bearer token and authenticates the OpenAI-compatible calls on each endpoints. The boto3 Converse examples signal with SigV4 as a substitute, drawing in your extraordinary AWS credentials from the setting, a profile, or a task. Configure each in case you intend to make use of Converse alongside the OpenAI-compatible APIs.

Deal with a long-term API key as an exploration-only credential. For manufacturing, the Grok 4.3 launch publish recommends short-term bearer tokens generated out of your IAM credentials with the aws-bedrock-token-generator package deal, as a result of they expire mechanically and maintain entry tied to your IAM id, and that steerage applies equally right here.

Working with reasoning effort

Reasoning is energetic on Grok 4.6 by default, and also you configure how a lot of it the mannequin spends by the reasoning parameter with low (the default), medium, excessive, or xhigh. The xhigh stage is new relative to what the Grok 4.3 launch publish documented, the place the degrees had been none, low, medium, and excessive.

Reasoning content material is encrypted. You may have it returned by passing embody: ["reasoning.encrypted_content"] on a Responses API request, then ship that content material again on subsequent turns to present the mannequin its personal prior reasoning as context in a multi-turn dialog. The Chat Completions API doesn’t return reasoning tokens.

Encrypted reasoning is a Responses API characteristic, so this instance makes use of the OpenAI shopper relatively than the boto3 shopper from the Converse examples above:

from openai import OpenAI

shopper = OpenAI()  # OPENAI_BASE_URL factors on the bedrock-runtime endpoint

response = shopper.responses.create(
    mannequin="us.xai.grok-4.6",
    reasoning={"effort": "excessive"},
    embody=["reasoning.encrypted_content"],
    enter="Clarify quantum entanglement merely.",
)
print(response.output_text)

As a result of reasoning is by default and energy is per request, effort stage is an actual price and latency management. Run quick extraction and classification calls at low, and reserve excessive or xhigh for planning steps and lengthy agent trajectories the place an early mistake compounds. Benchmarking effort ranges in opposition to your individual workload is the quickest approach to discover the place larger reasoning stops incomes its token price.

Get began

Grok 4.6 on Amazon Bedrock offers you a mannequin xAI constructed for long-running brokers and impressive interactive work, with a 500K token context window, 4 reasoning effort ranges, picture enter, immediate caching, and a selection between the OpenAI-compatible bedrock-mantle endpoint and the bedrock-runtime endpoint with Converse API and cross-Area inference assist.

To begin constructing, evaluate the Grok 4.6 mannequin card for the present Area record, characteristic matrix, and parameter particulars, and test the Amazon Bedrock pricing web page for token charges. For those who generated a long-term Amazon Bedrock API key for exploration, delete it from the Amazon Bedrock console when you find yourself completed. A standing credential you now not want solely widens your account’s publicity floor.

Sources

In regards to the authors