Grok 4.6 lands on Bedrock at $2/$6, turning an xAI decision into a line item

xAI made Grok 4.6 generally available on Amazon Bedrock on Wednesday, seven days after the model's own release ^[1], in an announcement dated August 19 ^[2]. That collapses the usual distance between a model launch and an enterprise deployment, because AWS customers can buy it through a purchasing and deployment path they already have rather than arranging separate vendor onboarding and another cloud integration ^[10].

The rate card is short. Bedrock access is priced at $2 per million input tokens and $6 per million output tokens, matching the rates in xAI's own launch materials ^[3]. The model carries a 500,000-token context window and four configurable reasoning efforts: low, medium, high and xhigh ^[4]. Output is billed at three times input ^[18], and filling the entire context once as input costs one dollar ^[11], so on any agent workload the prompt is not the expensive part.

The announcement lists one price, not four. Reasoning effort is a volume lever rather than a rate lever: higher effort can improve performance on multi-step work while producing more billable output tokens ^[5]. The gap between a low-effort call and an xhigh call therefore arrives on the invoice as token count, not as a different unit price ^[13]. For a team already spending inside Bedrock, that reframes the question as which effort setting a workload actually needs, compared against the other models in a catalog AWS markets on the ability to evaluate and switch providers ^[12].

The context window moved the other way. According to runtimewire.com, 500K is smaller than the 1 million tokens previously listed for Grok 4.3 on Bedrock ^[14], a halving ^[19], and the publication frames the open question as whether Grok 4.6's agentic performance earns its higher price against both Grok 4.3 and rival models already sold through Amazon ^[6]. Long-context retrieval work and long-horizon agent work are now two different purchases from the same vendor.

On the evidence side, AWS attributes the performance claim to the model's maker: according to SpaceXAI, Grok 4.6 reaches frontier intelligence across several agentic coding and knowledge work benchmarks and matches other frontier models specialised for coding ^[20]. The Bedrock announcement carries no separate AWS evaluation, latency measurements or production reliability data ^[7]. AWS says the model is available in all Regions where Bedrock is offered, with cross-Region inference, monitoring and logging, and enterprise-grade security and privacy ^[8], while runtimewire.com notes the announcement points developers to documentation without naming exact Regions or interfaces ^[9].

What xAI says it built: an August 12 model announcement described a longer supplemental training run than Grok 4.5, using model-generated reasoning data, engineering data, supervised fine-tuning trajectories regenerated by Grok 4.5, and reinforcement-learning tasks spanning coding, knowledge work, web development, computer-aided design and kernel optimisation ^[15]. xAI also says it observed more self-testing and verification during longer tasks, which are its own findings from its own testing ^[21]. AWS lists the model under the SpaceXAI name ^[16], following a $20 billion Series E announced on January 6 and SpaceX's acquisition of xAI effective February 2, per SpaceX offering documents ^[17].

Worth watching: whether any independent latency or reliability numbers appear for the xhigh tier, since that is where the billing risk sits ^[7][5]; whether Grok 4.3 stays the cheaper choice for jobs that need the full million tokens ^[14]; and whether AWS publishes Region-level detail beyond the blanket availability statement ^[8][9].