One week after xAI released Grok 4.7, the model has crossed into enterprise distribution. AWS's Machine Learning Blog announced on Monday that the model is now available on Amazon Bedrock — a frontier model 'built for coding, long-running agents, and knowledge work,' with a 500K token context window and four configurable reasoning effort levels: low, medium, high, and xhigh.

The Bedrock listing matters because of who buys through that shelf. Bedrock is the managed platform where more than 100,000 organizations build generative AI applications, and each new frontier lab in the catalog makes the underlying model more interchangeable. For xAI, launched on September 21 at $2 per million input tokens and $6 per million output tokens, Bedrock is instant enterprise reach without building an enterprise sales motion.

Endurance over speed#

xAI positions Grok 4.7 as its most capable model for coding and knowledge work, and the theme is endurance rather than raw speed: the model works longer on difficult tasks and checks its own output more carefully before moving on. It uses a new, larger base model trained with a longer reinforcement-learning run over a harder mix of tasks — deliberately weighted toward problems that take many hours to complete. Two capabilities came out of that, per xAI: better self-verification, and more effective use of the 500K context window on long tasks.

For anyone building agents, the self-verification behavior is the detail worth noting: a model that checks its own work before continuing fails less catastrophically on long trajectories, where an early mistake otherwise compounds through every later step.

Illustration of a long document being checked by a neural mesh, representing self-verification.
Illustration generated for AI Frontier Post.

The reasoning dial — and its cost#

Reasoning is always active on Grok 4.7, and the effort level is the primary lever over cost and latency. You set it through the reasoning parameter on the Responses API or additionalModelRequestFields on Converse. The default is high — worth setting explicitly, since leaving it unset on latency-sensitive or high-volume calls spends more reasoning tokens than those calls need.

The tradeoff is real and measured. Citing Artificial Analysis' independent evaluations, the AWS post reports Grok 4.7 improving over Grok 4.6 across the suite — Intelligence Index 46 vs. 44, Coding Agent Index 56 vs. 47, AA-Briefcase long-horizon knowledge work 1,657 vs. 1,546 Elo, and hallucination rate down to 29% from 34%. But the gains come with roughly double the output tokens per task — about 81,000 vs. 38,000. The practical rule the AWS post suggests: short extraction and classification calls belong at low; multi-step planning and long agent trajectories are where high and xhigh earn their tokens.

How it's packaged on Bedrock#

Grok 4.7 accepts text and image input and returns text. It is served on the bedrock-runtime endpoint through cross-Region inference profiles — you name a profile, not a bare model ID: us.xai.grok-4.7 for the US geographic profile, which keeps processing inside the US for data-residency requirements, or global.xai.grok-4.7 for the Global profile, which routes to any supported commercial region, spreads load, and is priced below the geographic one. Three service tiers are offered: standard (pay-per-token, no commitment), priority (faster, prioritized processing for a premium), and flex (lower-cost access for work that isn't time-sensitive).

The model supports the Responses, Chat Completions, and Converse APIs. Because it is OpenAI-compatible, the OpenAI SDK works against the /openai/v1 path with a Bearer (a Bedrock API key or a short-term token minted from IAM credentials); the AWS SDKs reach the same model through Converse with ordinary AWS credentials. Standard Bedrock features attach too: implicit prompt caching for repeated prompt prefixes, Bedrock Guardrails (content filters, denied topics, PII redaction), structured outputs constrained to a JSON Schema, and invocation logging into CloudWatch with token counts including reasoning tokens — an audit trail for long agent runs.

Illustration of a wireframe globe with data routes between regions.
Illustration generated for AI Frontier Post.

Safety and operations#

On safety, xAI says Grok 4.7 was built with an entirely new safeguard stack and is the strongest model it has tested on refusals and jailbreak resistance — while still remaining useful for legitimate dual-use work in domains like cyber security. It has also begun giving selected cyber security partners invite-only access to the model's red-team capabilities for defense research.

Operationally, AWS flags the obvious things enterprise teams forget. Treat a long-term Bedrock API key as an exploration-only credential and delete it from the console when finished; use short-term Bearer generated from IAM credentials for production. And on permissions, bedrock:InvokeModel is evaluated against three resources — the account's default project, the named inference profile, and the underlying foundation model — so a policy naming us.xai.grok-4.7 does not cover global.xai.grok-4.7.

What to watch#

First, whether the reasoning dial gets used as a cost lever in practice: the double-token tradeoff means Grok 4.7's sticker price is not its real price at xhigh. Second, how fast enterprise adoption shows up — Bedrock listings are AWS's fastest distribution channel, but per-token pricing across the three service tiers (on the Bedrock pricing page) will decide who actually routes traffic. And third, the commoditization trend itself: every frontier lab added to Bedrock makes the model layer more interchangeable and the platform stickier — AWS doesn't need to out-benchmark anyone to win that game.