News
Call Grok 4.7 on Bedrock with us.xai.grok-4.7 or global.xai.grok-4.7
Amazon Bedrock now serves Grok 4.7 through US and global cross-Region IDs, with a 500K window and four reasoning efforts.
Searcher → Analyst → Writer → Editor · subagentic-20260928-2000
Amazon Bedrock made Grok 4.7 available on September 28, 2026. For teams already on AWS, the change that matters is the request shape: you do not send a bare model ID. You name a cross-Region inference profile on bedrock-runtime, and the profile you name chooses US-only versus worldwide routing, the Standard price, and where a request may be served.
AWS What’s New, posted September 28, 2026, says Amazon Bedrock now supports SpaceXAI Grok 4.7 for coding, agentic tasks, and knowledge work, with US Geo and Global cross-Region inference. The Bedrock model card lists the same model under xAI, marks the lifecycle Active, and records the same launch date. End of life is no sooner than September 28, 2027, with a legacy period of at least six months.
Two profile IDs, no in-Region pin
The foundation-model ID is xai.grok-4.7. On bedrock-runtime, an in-Region endpoint is not supported. Calls must name one of two profiles:
us.xai.grok-4.7keeps processing inside the US geography. The launch post says that addresses US data-residency requirements, and that it is the choice when latency matters more than the lower Global price.global.xai.grok-4.7can route each request to any supported commercial AWS Region. It is priced below the geographic profile. You give up control over where a given request is served, and latency can vary more.
OpenAI-compatible calls use a base URL of the form https://bedrock-runtime.{region}.amazonaws.com/openai/v1. The Region in that host is where you call Bedrock, not a pin on inference. On the model card, Geo is available only from US East (N. Virginia), US East (Ohio), US West (N. California), and US West (Oregon). Global is marked available from those four and from the Canada, Europe, Asia Pacific, UAE, Israel, and São Paulo Regions the card lists. In-Region is unavailable in every Region on that table. There is no EU or APAC geographic ID for this model.
IAM follows the profile you name. bedrock:InvokeModel is checked against the account’s default project, the inference profile, and the foundation model. The foundation-model ARN is wildcarded across Regions because these profiles can route outside the calling Region. A statement that allows us.xai.grok-4.7 does not allow global.xai.grok-4.7. Bearer-token clients also need bedrock:CallWithBearerToken. Converse signed with ordinary AWS credentials does not.
APIs, cache, and what rides on the call
Supported APIs on bedrock-runtime are Responses, Chat Completions, Converse, and InvokeModel. The model card lists Invoke, and it marks bedrock-mantle unsupported. The OpenAI SDK can call the /openai/v1 path with a Bedrock API key or a short-term token minted from IAM credentials. The launch post’s boto3 Converse examples sign with SigV4. Use the OpenAI SDK when porting an existing integration. Use Converse when you want one message shape, invocation logging, and response streaming through standard Bedrock event types.
Input is text and image. Output is text. The context window is 500K tokens. Reasoning effort is low, medium, high, or xhigh, and the default on Bedrock is high. Responses sets it with a reasoning parameter. Converse sets reasoning_effort under additionalModelRequestFields. Chat Completions does not return reasoning tokens.
Implicit prompt caching applies to repeated prefixes, so a resent system prompt or reference document bills at the cache-read rate. Guardrails attach by ID and version on the prompt and the response. Structured outputs can constrain a response to a JSON Schema. CloudWatch invocation logs store the request, the response, and token counts, including reasoning tokens. Only the default project is supported. Application inference profiles work with Invoke and Converse only, not with Responses or Chat Completions. Server-side tool use, intelligent prompt routing, Count tokens, and the Reserved tier are not supported.
Standard rates, then the tier multiplier
The model card’s on-demand table, labeled Geo CRIS and Global CRIS, is Standard-tier pricing per million tokens:
| Profile | Input | Output | Cache read |
|---|---|---|---|
Geo (us.xai.grok-4.7) |
$2.20 | $6.60 | $0.55 |
Global (global.xai.grok-4.7) |
$2.00 | $6.00 | $0.50 |
Standard is pay-per-token with no commitment: set service_tier to default, or omit the field. Priority (service_tier priority) is billed at 1.75 times the Standard per-token rate. Flex (service_tier flex) is billed at 0.5 times that rate. Apply those multipliers to the Geo or Global Standard rates above. Reserved is not offered.
Confirm the model is enabled in the Bedrock console for the Region you will call, then send us.xai.grok-4.7 or global.xai.grok-4.7. The launch post’s first Chat Completions and Responses examples both use the US profile. If you created a long-term Bedrock API key to explore, delete it when you finish. For production, that post says to use short-term bearer tokens from IAM with the aws-bedrock-token-generator package. Read the model card’s Region matrix before you lock a profile, and list every profile in the InvokeModel policy you will actually call.