On 28 September 2026 AWS’s machine-learning blog confirmed Grok 4.7’s arrival in the Amazon Bedrock catalogue. It is not the model’s debut—xAI introduced it on 21 September—but its enterprise packaging on Amazon’s cloud, with profile IDs `us.xai.grok-4.7` and `global.xai.grok-4.7` rather than a bare model ID.
xAI positions Grok 4.7 as its most capable model for coding and knowledge work, with an emphasis on endurance: it works longer on hard tasks and verifies its own output before moving on. It trains on a larger base and a longer RL run weighted toward many-hour problems. xAI also says it natively understands the Grok Bot harness.
Artificial Analysis, as an independent reference cited by AWS, places the Intelligence Index at 46 versus 44 for Grok 4.6, the Coding Agent Index at 56 versus 47, and Elo gains on AA-Briefcase and GDPval-AA. The trade-off is explicit: ~81k output tokens per index task versus ~38k for the predecessor. More quality, higher bills if effort is left on default.
On Bedrock, reasoning effort is set to low, medium, high or xhigh; the default is high. AWS warns teams to set it explicitly: inheriting high on high-volume classification or extraction burns reasoning tokens unnecessarily. Reasoning content is encrypted; it can be returned on Responses with `include: ["reasoning.encrypted_content"]` for multi-turn use.
The model accepts text and image and returns text. Global routing (more capacity, less residency control, lower price) and a U.S. Geo profile (U.S. residency) are available. Standard, Priority and Flex service tiers balance latency and cost. Implicit prompt caching, Guardrails by ID, JSON Schema structured outputs and CloudWatch invocation logging complete the operational pack.
OpenAI SDK compatibility against the `/openai/v1` path eases porting existing integrations; Converse unifies the message shape with other models in the account. AWS stresses that long-term API keys are for exploration only and that production should use short-lived tokens tied to IAM.
The announcement lands days after Neural Pulse covered Grok 4.7’s delay for extra RL: now the model not only exists but is purchased with an AWS bill, compliance and logging. For teams that already standardised on Bedrock—and that on 30 September saw OpenAI Managed Agents—Grok adds another column to the multi-model router.
On safety, xAI claims an entirely new safeguard stack that is strong on jailbreak resistance, with invite-only access to red-team capabilities for cyber-defence partners. That rhymes with Gemini 4 Argon’s Fairwind and Anthropic’s cyber programmes: the industry is segmenting dual-use capability behind allowlists.
Buyer risks include the AA-Omniscience hallucination rate still at 29% (better than 34% prior, but not low); double the output tokens; and dependence on cross-Region profiles that complicate IAM policies (each inference profile must be listed). A policy that names only `us.` does not cover `global.`.
Competitively, Bedrock strengthens the thesis that xAI—now under the SpaceXAI umbrella in much coverage—needs cloud channels for enterprise revenue beyond the Grok app. Microsoft had already moved Grok toward Copilot in September; Amazon answers with the classic foundation-model catalogue.
For long-running agents, AWS highlights self-verification as an antidote to early errors that compound. That holds only if the customer harness propagates encrypted reasoning and does not blindly drop effort to low for savings. Cost engineering becomes part of the prompt stack.
Neural Pulse sees Grok 4.7 on Bedrock as the industrialisation of the model after September’s relaunch. Success will not be another leaderboard, but whether teams choose `xhigh` only when agent ROI justifies it—and whether invite-only cyber safeguards scale without creating a grey market for access.
