Grok 4.7 Arrives on Amazon Bedrock with 500K Context Window and Configurable Reasoning

xAI's Grok 4.7 is now available on Amazon Bedrock, giving AWS customers direct access to the model through cross-Region inference profiles rather than a standalone deployment. The model ships with a 500K token context window and four configurable reasoning effort levels: low, medium, high, and xhigh.

Grok 4.7 was originally announced by xAI on September 21, 2026, in a launch xAI describes as its most capable model yet for coding and knowledge work. According to xAI, the model was trained on a new and larger base model, with a longer reinforcement learning run over a harder mix of tasks weighted toward problems requiring many hours to complete. xAI says this produced two effects: improved self-verification of its own output, and more effective use of the extended context window across long tasks.

Independent benchmark data

Artificial Analysis, which runs its own evaluations independent of vendor-reported figures, measured Grok 4.7 against its predecessor Grok 4.6 at xhigh reasoning effort:

Measure	Grok 4.7	Grok 4.6
Intelligence Index	46	44
Coding Agent Index	56	47
AA-Briefcase (long-horizon work, Elo)	1,657	1,546
GDPval-AA (professional work, Elo)	1,695	1,605
AA-Omniscience Index	32	30
AA-Omniscience hallucination rate	29%	34%
Output tokens per Intelligence Index task	~81k	~38k

The largest gains appear in long-horizon agentic knowledge work and coding agents run in xAI's own harness. The tradeoff: Grok 4.7 uses roughly double the output tokens per task compared to Grok 4.6, making reasoning effort a cost lever that developers need to set deliberately rather than accept by default.

How it's packaged on Bedrock

Grok 4.7 accepts text and image input and returns text output. It is served on the bedrock-runtime endpoint through two cross-Region inference profiles rather than a single model ID:

• Global: global.xai.grok-4.7 — routes requests to any supported commercial AWS Region, priced below the geographic option, with more variable latency
• US Geo: us.xai.grok-4.7 — keeps processing within the US, addressing data residency requirements

The model supports the Responses API, Chat Completions API, InvokeModel, and Converse API. Because it is OpenAI-compatible, developers can use the OpenAI SDK against the /openai/v1 path with a bearer token (a Bedrock API key or short-term IAM-derived token), or use AWS SDKs through Converse with standard AWS credential signing.

Bedrock-specific features available with Grok 4.7 include implicit prompt caching for repeated prompt prefixes, Amazon Bedrock Guardrails for content filtering and PII redaction, structured outputs constrained to a JSON Schema, and invocation logging through CloudWatch that captures token counts including reasoning tokens.

Pricing tiers

Three service tiers are available: Standard (pay-per-token, default), Priority (faster processing at a premium), and Flex (lower cost for non-time-sensitive work). Amazon states that per-token pricing across tiers is detailed on the Bedrock pricing page; pricing not yet disclosed in the source material.

Safety claims

According to xAI, Grok 4.7 was built with a new safeguard stack and is the company's strongest model tested on refusal and jailbreak resistance. xAI also states the model blocks only a small fraction of risky dual-use cyber security prompts while rarely blocking legitimate security research, and has begun giving select cyber security partners invite-only access to the model's red-team capabilities.

What this means

This is a distribution event, not a model launch — Grok 4.7 itself debuted via xAI's own announcement in September 2026. What's notable here is packaging: Bedrock customers get Grok 4.7 through the same Converse API and Guardrails infrastructure they already use for Anthropic, Meta, and other Bedrock models, lowering the switching cost for enterprises already standardized on AWS tooling. The two-tier inference profile structure — cheaper Global routing versus residency-guaranteed US routing — mirrors a pattern AWS has used for other frontier models, giving buyers an explicit cost-versus-control dial. The doubling of output tokens per task at higher reasoning effort is the detail worth budgeting for: teams that flip to xhigh reasoning without adjusting expectations will see both better benchmark performance and roughly double the token spend.