AWS Adds Grok 4.7 to Bedrock With Global and US Model IDs

Grok 4.7 arrived on Bedrock on September 28

Amazon Web Services announced Grok 4.7 availability on Amazon Bedrock on September 28, 2026. SpaceXAI had introduced the model on September 21 and made it available through the Grok API and partner platforms. The Bedrock release is new access to an existing model, not a second Grok 4.7 model launch.

SpaceXAI positions Grok 4.7 for coding, long-running agents and knowledge work. It says the model uses a new, larger base model than Grok 4.6, a longer reinforcement-learning run and a harder task mix weighted toward work that takes hours. Those are vendor descriptions. They do not establish how long the model can operate reliably in another tool harness or production environment.

Bedrock exposes two cross-Region model IDs

Bedrock requests use the geographic profile us.xai.grok-4.7 or the Global profile global.xai.grok-4.7. AWS documents xai.grok-4.7 as the underlying foundation-model ID but does not support an in-Region endpoint for it. Applications therefore need permission for the chosen inference profile and the underlying model resource.

AWS says the US profile keeps processing within the US geography. Global CRIS can route a request to any supported commercial AWS Region, increasing the available capacity and carrying the lower list price, but reducing control over the processing location and potentially adding latency variation. Teams with residency obligations should not translate “US geography” or “Global” into a contractual conclusion without checking their account, agreement, logs and current regional documentation.

The API surface is broad, but response handling differs

AWS documents support for the Responses API, Chat Completions API, InvokeModel and Converse API. The OpenAI-compatible APIs use the bedrock-runtime /openai/v1 path with a Bedrock API key or short-lived bearer token. Converse uses ordinary AWS SigV4 credentials. A deployment that uses both paths needs both authentication and IAM behavior covered in its tests.

Reasoning is active, and the documented effort levels are low, medium, high and xhigh, with high as the default. Responses and Converse express that control in different request fields. AWS also warns that Converse places reasoning in an earlier content block and the answer in a later text block. Code that assumes content[0] is always the final answer can therefore expose reasoning or return the wrong content.

Global Standard pricing matches the direct API headline

The Bedrock model card lists Global Standard rates per million tokens of $2 for input, $6 for output and $0.50 for cache reads. The US geographic profile is $2.20 input, $6.60 output and $0.55 cache read. SpaceXAI’s direct API documentation lists the model at $2 input and $6 output per million tokens; provider features, routing, caching and account terms still make these separate procurement choices.

Bedrock also supports Priority at 1.75 times the Standard per-token rate and Flex at 0.5 times Standard. Standard has no commitment, Priority buys prioritized processing, and Flex is intended for non-time-sensitive work. Reasoning effort can dominate the bill: AWS’s summary of Artificial Analysis notes that its xhigh Grok 4.7 run used roughly twice the output tokens per Intelligence Index task as Grok 4.6. Price a completed accepted task, including reasoning, retries and tool turns, rather than multiplying one prompt by a headline token rate.

A 500K context is a capacity limit, not a reliability result

Both providers document a 500,000-token context window, with text and image input and text output. AWS adds implicit prompt caching, structured outputs, Bedrock Guardrails and invocation logging. Those features can reduce repeated-prefix cost, constrain response shape and improve observability, but none proves accurate retrieval across the full context or prevents a long agent trajectory from compounding an early mistake.

Test context length progressively with evidence-bearing tasks from the intended workload. Measure whether required facts survive placement changes, distractors, tool-result growth and summarization. Set a maximum prompt size, reasoning effort, output budget, tool-call count and wall-clock duration before the trial. A large advertised window is not a reason to remove those bounds.

Benchmark and safety claims need separate labels

SpaceXAI reports gains on CursorBench 4.0, DeepSWE v1.1, Terminal-Bench 4.0, AA Briefcase, EEBench, legal and clinical evaluations. It reports 46.3% on CursorBench at xhigh effort and describes the model as stronger at verification and long-context management. These are provider-run, provider-selected claims with settings and footnotes that must travel with the numbers; they are not AccessAllGPT measurements.

Artificial Analysis independently reports a higher Intelligence Index, Coding Agent Index and long-horizon knowledge-work score than Grok 4.6, alongside substantially greater output-token use. That is useful secondary evidence but still does not predict a private workload. SpaceXAI also reports a new safeguard stack and low risky-prompt allowance on its own HackerBench evaluation. Treat that as a vendor safety claim, not authorization for cyber work or proof that tool-enabled agents are safe.

Run a profile-specific pilot before production

Start with Global Standard for a reversible, non-sensitive evaluation when cross-Region processing is acceptable; use the US profile when US-geography processing is a requirement worth the price difference. Pin the profile ID, API surface and reasoning effort. Replay complete tasks against the current model, capturing accepted-task quality, latency, input, cache-read, reasoning and output tokens, tool calls, retries, refusals, malformed responses and cost.

Move to Priority only when measured latency has business value, or Flex when delayed completion is acceptable and retry behavior is understood. Wait when account access, routing, residency, logging or price evidence is missing. Reject unattended consequential actions without deterministic authorization, bounded permissions, approval and rollback. The immediate next step is a small shadow evaluation—not a global model-string replacement.

Disclosures

AccessAllGPT did not receive SpaceXAI, xAI, AWS or Artificial Analysis credentials, credits, early access, a briefing, benchmark data, review or compensation for this article. None of the cited organizations sponsored, reviewed or endorsed it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with SpaceXAI, xAI, Amazon Web Services or Artificial Analysis. Publication-wide relationships are listed on the disclosures page.