2026-09-28 Changes confirmed medium confidence

AWS Adds Grok 4.7 to Amazon Bedrock With US and Global Profiles

The xAI model is available through four Bedrock APIs and two cross-Region profiles, giving teams new deployment choices while leaving performance and safety claims to independent testing.

Edited by Tyronne Panaino

Amazon Web Services published Grok 4.7 availability for Amazon Bedrock on September 28, adding xAI's model to a managed cloud catalog used by developers and enterprise platform teams. The release matters because it gives Bedrock customers a new route to the model through familiar APIs, regional controls and operational tooling, without establishing that Grok 4.7 will be the best choice for any particular workload.

What changed in Bedrock

The official AWS launch post says Grok 4.7 is now available for coding, long-running agents and knowledge work. AWS lists a 500K-token context window and four configurable reasoning-effort levels: low, medium, high and xhigh. Those settings give teams a direct way to test how additional model effort affects their own latency, token use and output quality instead of treating one default as suitable for every task.

Grok 4.7 accepts text and image input and returns text. Bedrock supports it through the Responses API, Chat Completions API, InvokeModel and Converse API. That breadth matters for adoption: teams can evaluate the same model through an OpenAI-compatible interface or through AWS-native request patterns, depending on the software and controls they already use.

AWS describes the model as better suited to long work and self-verification, but those capability claims originate with xAI. The supported release fact is narrower: the model and these interfaces are available in Bedrock.

US and global profiles create a deployment choice

AWS serves Grok 4.7 through cross-Region inference profiles rather than a bare model identifier. The US profile, `us.xai.grok-4.7`, keeps processing within the US geography. The global profile, `global.xai.grok-4.7`, can route requests across supported commercial AWS Regions.

That distinction turns model selection into an infrastructure decision as well as an evaluation decision. A team with data-residency or internal routing requirements should verify the US profile against its own policy. A team considering the global profile should test where the service is available, how routing affects latency and whether its governance rules allow that processing pattern. The launch page provides the profile boundaries, but it does not replace an organization's architecture and compliance review.

Bedrock controls can wrap the model

AWS documents compatibility with implicit prompt caching, Bedrock Guardrails, structured outputs and invocation logging. Prompt caching can reuse repeated prefixes. Guardrails can apply configured filters and redaction policies to requests and responses. Structured outputs can constrain a response to a JSON Schema, while invocation logging can send request, response and token information to Amazon CloudWatch.

These are platform controls around a model, not proof of the model's inherent reliability or safety. Logging improves the evidence available for review; a schema can make output easier to parse; guardrails can enforce configured policies. None of those mechanisms guarantees that an agent will choose the right action, remain accurate over a long task or behave consistently under adversarial input. Teams still need task-specific evaluations, access limits, human escalation paths and monitoring that measures real failure modes.

Evidence quality and the next checkpoints

The fetched evidence is one official AWS launch article, so confidence is medium. It establishes the Bedrock availability, interfaces, profile identifiers and documented platform integrations. It does not independently reproduce xAI's performance, training or safeguard claims, and no independent evaluation was fetched for this release.

The next useful checkpoints are measured results from production-like workloads: completion quality at each reasoning level, total token use, latency, failure recovery, image-input behavior and regional availability. Those results would show whether Bedrock distribution changes only procurement and integration, or materially changes how teams can operate Grok 4.7 in governed systems.

Status

Confirmed. AWS's official post supports the product availability and configuration details. Internal confidence is medium because the source is first-party and the model-performance and safety claims were not independently validated in the fetched evidence.