The model dropped a week ago. The distribution news is today.
xAI announced on Wednesday that Grok 4.6 is generally available on Amazon Bedrock, with a 500K token context window, configurable reasoning effort from low to xhigh, and pricing at $2 per million input tokens and $6 per million output tokens. The model itself launched on August 12, built on Grok 4.5 with a longer supplemental training run and a focus, in xAI’s words, on long-running agents and ambitious interactive and visual work.
The Bedrock addition is the part that changes the deal.
The shape of the long-running agent
Look at the spec as a workload description. A 500K context window is not for chat. It is for an agent that has to hold a codebase, a research trail, and its own working history in one pass. Configurable reasoning effort is an admission that not every step in a long agent run deserves the same compute, which is how you keep a ten-hour job from costing a small fortune.
xAI says the model matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index and is available in Cursor, Grok Build, the API, and partners like OpenRouter, Vercel, and Cloudflare. On Bedrock it comes with cross-region inference profiles, a US geo profile for data residency and a global one, and the standard Responses, Chat Completions, and Converse APIs.
Distribution is the product
I wrote about OpenAI landing on Bedrock in May, and the point was that enterprise AI adoption is blocked less by model intelligence than by trust paths: identity, logging, spend controls, vendor risk, the existing cloud architecture. A frontier model that shows up inside the cloud the company already runs has a shorter road to production than one that asks for a new operational relationship.
Grok 4.6 on Bedrock is the same move from the other side of the table. xAI does not need to win the enterprise procurement battle model by model. It needs the model to be one line in the same Bedrock console where the customer already manages the rest. When the frontier models are interchangeable behind one API, the decision stops being which lab is best and starts being which one is cheapest, fastest, and easiest to govern for the specific workload.
That is where the agent war actually is now. Not on the leaderboard. In the console.