The Pulse
Amazon Bedrock Adds xAI’s Grok 4.6 for 500K-Token Agents
Amazon Bedrock now offers xAI’s Grok 4.6 for coding, knowledge work and long-running agents. The model brings a 500,000-token context window, four reasoning levels, image input and access through AWS runtime APIs.
AI.info Team ·
A 500,000-token context window is the headline specification as Amazon Bedrock adds xAI’s Grok 4.6 to its model catalog. AWS says the model is designed for long-running agents, coding and knowledge work, with four configurable reasoning levels and support for text and image input.
AWS announced the availability on September 21, 2026. The company says Grok 4.6 launched on Bedrock on August 18, making it xAI’s second model on the service after Grok 4.3.
The release gives developers another option for agents that must retain large amounts of context while working through multi-step tasks. Grok 4.6 is available through both Amazon Bedrock’s bedrock-mantle endpoint and the bedrock-runtime endpoint, extending access beyond the OpenAI-compatible Mantle interface used for Grok 4.3.
Grok 4.6 targets work that continues across many steps
xAI positions Grok 4.6 as an upgrade from Grok 4.5 for tasks that require sustained execution rather than a single answer. AWS describes use cases including research, information analysis, software development and the creation of interactive applications or other work products.
The model supports low, medium, high and xhigh reasoning effort. Developers can choose a lower setting for shorter extraction or classification jobs, or spend more computation on planning and longer agent trajectories. AWS says reasoning is active by default and can be configured per request.
xAI’s training description, reproduced in the AWS announcement, emphasizes agentic reinforcement-learning tasks across knowledge work, general coding, kernel optimization, web development and computer-aided design. The company also says Grok 4.6 showed more self-testing and verification during longer task sequences.
Bedrock adds Converse and cross-Region access
Grok 4.6 supports the Responses API, Chat Completions API and Amazon Bedrock’s Converse API. Converse support gives developers a common message format for applications that may switch between models available through Bedrock, while streaming is available through both converse and converse_stream.
The bedrock-runtime endpoint uses cross-Region inference profiles rather than a bare model identifier. Applications can call us.xai.grok-4.6 for US geographic routing or global.xai.grok-4.6 when the widest capacity pool is preferred. AWS documentation lists geographic availability in US Regions including Northern Virginia, Ohio, Northern California and Oregon, with global routing available across a broader set of locations.
Grok 4.6 is not available for in-Region inference through bedrock-runtime. In-Region access is available through bedrock-mantle in US West, according to the Amazon Bedrock model card.
Tools, images and logging shape the agent workflow
The model accepts image input and returns text. AWS lists tool calling, structured output, response streaming, prompt caching and encrypted reasoning content among the supported capabilities, although the feature set differs between the two endpoints.
Bedrock-mantle supports structured outputs and client-side tool calling. Bedrock-runtime supports the Converse API, invocation logging and cross-Region inference, but AWS says structured outputs and server-side tool use are not available there. Developers therefore face a practical endpoint choice: Mantle provides more OpenAI-compatible application features, while runtime fits workloads that depend on AWS’s standard Bedrock interface and operational logging.
Invocation logging can capture request and response bodies, token counts, reasoning-token usage and the inference profile used. For agents operating without continuous human supervision, those records provide a way to inspect what the system received and returned during a run.
Pricing favors global routing when residency permits
AWS says the global inference profile costs $2 per million input tokens, compared with $2.20 for the US geographic profile. The difference applies to standard input pricing and may matter for agents that repeatedly submit large prompts or documents.
Grok 4.6 also supports Standard, Priority and Flex service tiers. Priority is priced at 1.75 times the Standard rate, while Flex is billed at half the Standard rate for workloads that can tolerate less predictable processing times. Prompt caching can reduce the cost of repeated input prefixes, a useful option for agents that resend a large system prompt or reference document across multiple turns.
The release puts xAI’s newest model inside the same managed service as models from other providers, but the operational trade-offs are specific. Developers must choose between endpoints, account for cross-Region routing, set reasoning effort against cost and latency, and decide how much of an agent’s work should be retained in logs.
For teams testing Grok 4.6, AWS directs users to the model card for regional availability and API details. The model identifier is xai.grok-4.6 on bedrock-mantle and us.xai.grok-4.6 or global.xai.grok-4.6 on bedrock-runtime.