Grok 4.7 Lands on Amazon Bedrock With a 500K-Token Context Window — and a Benchmark Bill of 81,000 Output Tokens Per Task

Amazon Web Services published its launch post for Grok 4.7 on Amazon Bedrock on September 28, putting xAI's frontier model into the managed cloud catalog just a week after xAI shipped it on September 21. The model is available for coding, long-running agents and knowledge work, with a 500K-token context window, text-and-image input, and support for tool calling.

Access runs through cross-region inference profiles rather than a bare model ID: us.xai.grok-4.7 keeps processing inside the US geography, while global.xai.grok-4.7 can route requests to any supported commercial AWS Region — a distinction that turns model selection into a data-residency decision. The model is served through four APIs (Responses, Chat Completions, InvokeModel and Converse), including an OpenAI-compatible endpoint, so teams can call it with an OpenAI SDK or with native AWS credentials without ever opening an xAI account.

A distinctive control is the reasoning-effort setting, with four levels — low, medium, high and xhigh. AWS suggests lower effort for short extraction or classification tasks and higher effort for complex planning or long agent runs; the setting directly affects cost and latency. Bedrock wraps the model with implicit prompt caching, Guardrails, structured outputs and invocation logging.

The performance claims in the launch post come from third-party benchmark firm Artificial Analysis, and they carry a caveat AWS itself notes: the two model versions were evaluated at different reasoning-effort conditions. Grok 4.7 scores an Intelligence Index of 46 versus 44 for Grok 4.6, a Coding Agent Index of 56 versus 47, and an AA-Briefcase Elo of 1,657 versus 1,546. The number buyers should model before migrating is output volume: average output per Intelligence Index task rose to roughly 81,000 tokens from about 38,000 — consistent with a model tuned to grind through long tasks rather than sprint, but a cost line that can double a bill.

Pricing differs by channel. On xAI's own API, Grok 4.7 costs \$2 per million input tokens and \$6 per million output tokens below a 200K prompt window (with cached input at \$0.50), rising to \$4 and \$12 above it — figures the company reports itself. The Bedrock launch post does not list prices, pointing customers to the Bedrock pricing page and console.

xAI also makes safety claims for the model that are, by nature, company-reported: it says Grok 4.7 was built with an entirely new safeguard stack, tops LatchBio's biosafety benchmark at 62.4 percent, and allows only 3.3 percent of risky dual-use cyber prompts through on its HackerBench v0.3 while rarely blocking legitimate security work. Select cybersecurity partners are getting invite-only access to its red-team capabilities.

Strategically, the listing completes a notable shift: Grok, born as the in-house model of a social platform, is now an enterprise catalog entry that a fintech engineer can swap in by changing one environment variable. It sits in a Bedrock catalog that added Anthropic's Claude Sonnet 5.5 on September 29 — including a GovCloud launch — and Grok 4.6 in August, meaning AWS now happily distributes models from two rivals to the same customers.

The trade is deliberate on both sides. xAI gets enterprise reach it could never build alone; AWS reinforces its position as the neutral multi-model hub. The pressure that creates lands on differentiation: when frontier weights are one line of configuration away, the durable moat moves up the stack — into harnesses, agents and the economics of how many tokens a task really takes.