xAI’s Grok 4.6 reaches Amazon Bedrock with two deployment paths

Endpoint choice determines available features

Bedrock Mantle exposes the model as xai.grok-4.6 through an OpenAI-compatible endpoint. It supports Chat Completions and Responses, as well as client-side tool calling, JSON Schema structured output, reasoning, prompt caching, streaming, projects and abuse detection. In-Region inference is available through this path in US West (Oregon), identified as us-west-2.

Bedrock Runtime uses the cross-Region profiles us.xai.grok-4.6 and global.xai.grok-4.6. It supports OpenAI-compatible calls plus Converse and Converse Stream, allowing applications to use Bedrock’s common message format and streaming events. Tool calling works through either Converse’s toolConfig or the OpenAI-compatible tools parameter.

The two paths are not feature-equivalent. Developers needing JSON Schema structured output must use Bedrock Mantle, while those needing Converse or invocation logging must use Bedrock Runtime. Runtime does not support the Invoke API, structured outputs, server-side tool use, intelligent prompt routing, token counting or application inference profiles. Its project support is limited to the default project.

Runtime adds Amazon Bedrock Guardrails across its APIs, including content filters, denied-topic rules, personally identifiable information redaction and word policies applied to prompts and responses. With model invocation logging enabled, Amazon CloudWatch records can capture request and response bodies, token counts including reasoning tokens, and the selected inference profile. Those controls can help operators review unattended agent runs, but they do not remove the need to test model behavior and policy settings against the intended workload.

Cross-Region inference brings residency and capacity tradeoffs

Bedrock Runtime does not offer in-Region inference for Grok 4.6. The US geographic profile routes within the United States for workloads with data-residency requirements, while the global profile can route worldwide to draw on a broader capacity pool. Geo inference can be initiated from us-east-1, us-east-2, us-west-1 and us-west-2; AWS says global inference is available from more than 30 Regions across multiple continents. AWS advises checking the model card before fixing a deployment Region because availability can change.

The source lists global Standard input at $2 per million tokens, compared with $2.20 for the US geographic option. Grok 4.6 also has Standard, Priority and Flex service tiers. Priority costs 1.75 times the Standard rate, while Flex costs half the Standard rate. Applied to the cited $2.20 Standard input price, those multipliers produce $3.85 for Priority and $1.10 for Flex per million input tokens.

xAI separately lists prices starting at $2 per million input tokens and $6 per million output tokens, with a fast variant at twice those prices. These figures are reference rates from the supplied source rather than permanent quotes; AWS explicitly directs customers to its pricing page for current charges.

Prompt caching can reduce repeated-input costs: AWS says cached input is billed at roughly one-quarter of the normal input rate. Caching applies to a repeated prefix, so applications that resend a large system prompt or document should place stable content first and verify the cached-token count before assuming the discount in a cost model.

Training claims and reported benchmark results

xAI says Grok 4.6 received a longer supplemental training run than Grok 4.5, using curated model-generated material for reasoning and technical concepts, engineering data, and an updated optimizer and training recipe. It used Grok 4.5 to regenerate supervised fine-tuning trajectories across reasoning levels, agent harnesses, STEM, software engineering and knowledge work, then filtered problematic traces with model-based checks. Agentic reinforcement learning covered general knowledge work and coding as well as kernel optimization, web development and computer-aided design.

According to xAI, longer task trajectories produced more self-testing and verification, while visual and interactive projects received stronger first passes. The company also says safeguards were improved and calibrated through its widest pre-deployment testing suite to date, followed by post-deployment and third-party testing. These behavioral and safety descriptions are vendor claims: the AWS article does not present detailed independent validation or results from the cited third-party safety testing.

For Grok 4.6 High, xAI reported an AA Intelligence Index score of 61, GDPVal-AA v2 of 1753, CursorBench v3.2 of 69.9%, DeepSWE v1.1 of 65.9%, FrontierCode v1.1 Extended of 61.3%, APEX-Agents of 57.5%, Terminal-Bench v3.0 of 26%, APEX-SWE of 56.4%, AA-Briefcase of 1577 and Harvey LAB (Vals) of 15.8%. The figures were published by xAI for its August 12, 2026 launch; they are reported results, not an independent comparative evaluation performed by AWS.

Artificial Analysis describes its Intelligence Index as a composite of nine evaluations covering areas including agentic tool use, reasoning, knowledge reliability, long-context reasoning and quantitative analysis of documents and spreadsheets. Its AA-Briefcase Elo combines rubric pass rate, analytical quality and presentation quality. Because reasoning effort changes token consumption and latency, developers still need workload-specific evaluation across all four settings rather than treating a high-effort benchmark result as representative of every deployment.

Authentication and reasoning require separate integration choices

OpenAI-compatible requests to either endpoint can use an Amazon Bedrock API key as a bearer token and require bedrock:CallWithBearerToken. Converse calls through the AWS SDK instead use SigV4 with ordinary AWS credentials. Applications using both paths must configure both authentication mechanisms. AWS recommends treating long-term API keys as exploration credentials and using short-term bearer tokens tied to IAM identities for production.

Runtime permissions must cover the account’s default project, the named inference profile and the underlying foundation model. The US and global profiles are authorized separately, so permission to call us.xai.grok-4.6 does not automatically cover global.xai.grok-4.6.

Encrypted reasoning content can be returned through the Responses API and supplied again in later turns; Chat Completions does not return reasoning tokens. Converse sets effort through additionalModelRequestFields. Its response should be searched for the text block rather than read from a fixed first position because reasoning may occupy an earlier block.

These interface differences make Grok 4.6’s Bedrock availability less a single uniform deployment than a choice between capabilities. Mantle fits applications that require structured output and in-Region inference in Oregon, while Runtime provides Converse, logging, Guardrails and cross-Region routing. The appropriate choice depends on residency, observability, output-format and authentication requirements as well as model quality and cost.

Source: AWS Machine Learning Blog.

Definition. Grok 4.6 on Amazon Bedrock is a 500K-token text-and-image model offered through two endpoints with different APIs, deployment options and operational features.

Key takeaways

FAQ

Sources