xAI's Grok 4.7 AI model is now on Amazon Bedrock: what AWS says it offers businesses and developers

What happened

On September 28, 2026, AWS announced on its official blog that Grok 4.7, developed by xAI, is now available on Amazon Bedrock and has been added to the Bedrock model catalog. Amazon Bedrock is AWS's cloud platform where businesses can use AI models from different companies. AWS describes Grok 4.7 as a frontier model built for coding, long-running AI agents and knowledge work. AI agents are programs that carry out multi-step tasks on their own. The AWS post states that xAI published its Grok 4.7 announcement on September 21, 2026.

According to the AWS post, Grok 4.7 has a 500K-token context window. A token is a small chunk of text, and AI services usually count usage and charges in tokens. The context window is how much text the model can consider at once. The model accepts text and images as input and produces text. It supports four configurable reasoning effort levels, which set how much work the model does before answering: low, medium, high and xhigh. AWS quotes xAI as calling Grok 4.7 its strongest model for coding and knowledge work. According to xAI, the model works longer on hard tasks and checks its own output more carefully before moving on to the next step.

Read the full description

Sources are collected, independently checked, then reviewed by Jev.

Capabilities and training claimed by xAI

The AWS post relays xAI's account of how the model was built. Grok 4.7 uses a newer, larger base model. It was trained for longer with reinforcement learning, a method that improves a model through trial and feedback, on harder tasks. That training was deliberately weighted toward problems that take hours to complete. xAI says this makes the model better at verifying its own work and at making good use of its 500K context. The model was also trained to natively understand the Grok Bot harness, the software setup the model runs in.

• Stronger document and presentation generation (according to xAI).
• Improvements in professional knowledge work such as that of lawyers, nurses and financial analysts (according to xAI).
• xAI's evaluations cover CursorBench, DeepSWE, Terminal-Bench, AA Briefcase, EEBench, Harvey Legal Agent Benchmark and HealthBench Professional. The AWS post does not list the scores and points readers to xAI's announcement instead.

Third-party benchmarks: gains and trade-offs

The AWS post cites data from evaluation firm Artificial Analysis comparing Grok 4.7 with the previous version, Grok 4.6. AWS notes that Grok 4.7 was measured at xhigh reasoning effort. In the table, Elo is a rating system in which a higher number is better. The hallucination rate measures how often a model confidently states false information, so a lower number is better. All figures are as reported in the AWS post.

AWS points out the trade-off in the last row: the gains come with roughly twice the output tokens per task. In practice, the same task may use more computing power and cost more. For that reason, AWS recommends choosing the reasoning effort level deliberately rather than simply keeping the default.

Safety and security measures

The AWS post relays xAI's safety claims. According to xAI, Grok 4.7 uses an all-new set of safeguards. xAI says it is the strongest model it has tested at refusing harmful requests and resisting jailbreaks, which are tricks meant to make a model ignore its safety rules. Some fields, such as cybersecurity, are dual-use: the same knowledge can be used to defend or to attack. In such fields, xAI says the model lets through only a small fraction of high-risk requests while rarely blocking legitimate security work. xAI has also begun giving selected security partners invitation-only access to Grok 4.7's red-teaming capabilities, which simulate attacks, for defensive research.

On Bedrock, AWS says the model can be paired with Amazon Bedrock Guardrails. Guardrails apply content filters, denied topics, masking of personally identifiable information and word policies to both requests and responses. Each call can also be logged in Amazon CloudWatch, including the request, the response and the reasoning token counts. These logs give an audit trail for long-running agent tasks.

How it is offered on Bedrock: regions and service tiers

According to the AWS post, Grok 4.7 is offered through cross-region inference profiles. These are routing settings that decide which AWS data-center regions handle each request, so the model is not pinned to a single region. The model works with the OpenAI SDK, using either an Amazon Bedrock API key or a short-term token generated from IAM (AWS account) credentials. It can also be used through the Converse interface of the AWS SDK. Other platform features include implicit prompt caching, which reuses repeated parts of a request, and structured outputs, which make responses follow a set JSON Schema format.

AWS lists three service tiers. Standard is pay-per-token with no commitment. Priority offers faster, prioritized processing at an additional cost. Flex costs less and suits work that is not time-sensitive. The AWS post does not list prices and points readers to the Amazon Bedrock pricing page.

What it means in practice for general readers

• Most everyday users will not use Bedrock directly. Any effect on them will come mainly from businesses or developers building Grok 4.7 into their own products and internal tools.
• Businesses already on AWS get one more model to choose from on the same platform, and they can keep using existing controls such as Guardrails and logging.
• Some businesses have data residency rules, meaning data must be processed in a certain place. AWS says the US profile keeps processing within the US. Businesses with residency needs in other regions must check those themselves.
• AWS notes that output tokens are about double those of the previous version. At high reasoning effort, costs and response times may rise, so it is worth checking whether that level is really needed before adopting it.

Browse the latest news in this topic

• AI news