Grok 4.6 Is Now Generally Available on Amazon Bedrock

xAI’s flagship model has a second major cloud home. Grok 4.6 is now generally available on Amazon Bedrock, the company announced on August 19, 2026, one week after the model’s initial release. On Bedrock the model is priced at $2 per million input tokens and $6 per million output tokens, with a 500,000-token context window and four configurable reasoning effort settings: low, medium, high, and xhigh.

AWS confirmed the listing the same day, stating that Grok 4.6 is available in every AWS Region where Amazon Bedrock is offered, and framing the addition around enterprise security, monitoring, and cross-Region inference. The model sits in the Bedrock catalog under the model ID xai.grok-4.6.

The Grok 4.6 model card in the Bedrock user guide shows a meaningfully wider deployment than the earlier Grok listing on the service. Grok 4.6 runs on both of Bedrock’s endpoints: bedrock-mantle, the inference engine that carried the first Grok release, and the standard bedrock-runtime endpoint, where it is reachable through Geo and Global cross-Region inference profiles across more than 30 AWS Regions. The mantle endpoint serves the model In-Region from us-west-2 (Oregon).

AWS lists Bedrock-side pricing at $2.20 per million input tokens and $6.60 per million output tokens for In-Region and Geo cross-Region requests, with Global cross-Region inference at $2.00 and $6.00, plus cached-read pricing of $0.55 per million tokens for In-Region and Geo requests and $0.50 for Global cross-Region inference. Modalities span text and image input with text-only output, per the model card.

Grok 4.6 on Amazon Bedrock by the Numbers

• Context window: 500,000 tokens, per xAI’s announcement
• Direct pricing: $2 per million input tokens, $6 per million output tokens
• Bedrock pricing: $2.20 / $6.60 per million tokens In-Region; $2.00 / $6.00 via Global cross-Region inference, per the AWS model card
• Artificial Analysis Intelligence Index: 61, up from 56 for Grok 4.5 High, per xAI’s launch-day eval table
• Regional footprint: more than 30 AWS Regions on the bedrock-runtime endpoint

What Grok 4.6 Changed From Grok 4.3 on Bedrock

This is the second Grok model to reach Amazon Bedrock. Grok 4.3 joined the catalog in June 2026 as xAI’s first listing as a Bedrock model provider, priced at $1.25 per million input tokens and $2.50 per million output tokens with a 1-million-token context window, served exclusively through the mantle endpoint.

Grok 4.6 reprices the line upward — $2 and $6 per million tokens against Grok 4.3’s $1.25 and $2.50 — while halving the stated context window to 500,000 tokens and adding a fourth reasoning tier, xhigh, above the previous high setting. Where Grok 4.3 reached Bedrock on June 17, 2026, Grok 4.6 made the jump in seven days. The wider bedrock-runtime support also moves Grok from a single-Region, single-endpoint deployment into Bedrock’s cross-Region inference system, which is how enterprises on AWS typically route production workloads.

What Grok 4.6 Ships With

Grok 4.6 launched on August 12, 2026 as xAI’s flagship for long-running agents and interactive and visual work, arriving first in Cursor and the company’s Grok Build environment. Unite.AI covered the launch and, two days later, the model’s expansion into GitHub Copilot across eight development surfaces.

In xAI’s launch-day evaluation table, Grok 4.6 High scores 61 on the Artificial Analysis Intelligence Index, a composite of nine benchmarks, against 56 for Grok 4.5 High — level with GPT-5.6 Sol Max and a point behind the table’s leader, Fable 5 Max, at 62. The company’s posted results show larger generation-over-generation gains on agentic suites: 65.9% versus 54% on DeepSWE v1.1, 57.5% versus 47.1% on APEX-Agents, and 26% versus 15.7% on Terminal-Bench v3.0. These are vendor-reported figures; xAI notes that third-party scores in the table are the best of self-reported or publicly available results.

The training recipe behind those numbers, as the company describes it, pairs a longer supplemental training run on curated model-generated reasoning and engineering data with supervised fine-tuning trajectories regenerated by Grok 4.5 across reasoning efforts and agent harnesses, followed by reinforcement learning on agentic tasks spanning knowledge work, general coding, kernel optimization, web development, and computer-aided design.

How Teams Reach Grok 4.6 on Bedrock

Developers call the model through the OpenAI SDK pointed at a Bedrock base URL, or through AWS’s own Converse API on the runtime endpoint. Reasoning is active by default at the low setting, and encrypted reasoning content can be returned and passed back across turns for multi-turn conversations, per the model card. Feature support splits by endpoint: server-side tool use is unsupported on the runtime side, while the mantle side carries client-side tool calling and abuse detection; prompt caching works on both.

For AWS, the listing extends a run of third-party frontier additions to the catalog: OpenAI’s Daybreak cyber defense models landed on Bedrock earlier this month, and the platform recently added built-in web search for grounded responses. For xAI, Bedrock distribution puts Grok 4.6 inside the procurement, identity, and compliance perimeter where enterprise AWS customers already run their inference.