xAI's Grok 4.3 Arrives on Amazon Bedrock, Running on a New 'Mantle' Inference Engine Reached Through an OpenAI-Compatible Endpoint
AWS added Grok 4.3 to Bedrock on June 15, making xAI a model provider on the platform and debuting a separate Mantle inference path.
Overview
Amazon Web Services added xAI’s Grok 4.3 model to Amazon Bedrock on June 15, 2026. According to AWS, “Today, AWS announces the availability of xAI’s Grok 4.3 model on Amazon Bedrock,” and “xAI joins Amazon Bedrock as a model provider, giving you even more choice as you build generative AI applications.” The launch makes xAI, by the account of Digital Applied, “the third major independent lab there alongside Anthropic and OpenAI.”
The addition continues a broadening of Bedrock’s model roster that The Machine Herald previously reported when OpenAI’s models and Codex landed on the platform in April. What distinguishes the Grok launch is not just a new vendor but a new way of serving the model: Grok 4.3 does not run on Bedrock’s standard runtime.
What We Know
The official Bedrock model card lists the model launch date as June 15, 2026, a 1M-token context window, and describes Grok 4.3 as “a reasoning-first model that offers always-on and configurable reasoning effort (none, low, medium, high).” The card explains the design choice directly: “Because reasoning is always active rather than optional, it behaves more consistently across multi-step agent loops than models that can skip thinking.”
The most notable technical wrinkle is the serving path. Per the model card, Grok 4.3 “runs on Mantle, a new inference engine in Amazon Bedrock designed for price performance, with support for tool calling, structured output, and response streaming.” The model is reached not through Bedrock’s usual bedrock-runtime endpoint but through a bedrock-mantle endpoint, and the card notes the model “is available on the openai/v1/responses path on the bedrock-mantle endpoint.” The supported APIs are Chat Completions and Responses — both OpenAI-compatible — while Bedrock’s own Converse and Invoke APIs are not supported for this model, according to the model card. The card publishes the model ID as xai.grok-4.3 and an endpoint URL of the form https://bedrock-mantle.{region}.api.aws/openai/v1.
At launch the model card lists in-region availability in three US regions: us-west-2 (Oregon), us-east-1 (N. Virginia), and us-east-2 (Ohio), with cross-region geo and global routing not yet supported, according to the model card. The same card documents that Grok 4.3 uses defaults that differ from the standard OpenAI specification — temperature defaults to 0.7, top_p to 0.95, and max_completion_tokens to 131072 — and that on the Chat Completions API the model does not return reasoning tokens.
On pricing, both Digital Applied and Basenor report on-demand rates of $1.25 per million input tokens and $2.50 per million output tokens, with cached input at $0.20 per million. Digital Applied adds a caveat that requests exceeding 200,000 total tokens are billed at a higher context tier. Basenor reports a maximum output of 30,000 tokens.
AWS positions the model for enterprise tasks. The official launch note says Grok 4.3 is “especially well suited to enterprise workloads such as customer support, web development, case law research, and financial document Q&A,” according to AWS. Basenor frames the appeal around “long-document workloads like contract review, case law research, and financial document analysis,” and cites a December 2025 study that placed Grok at an 8% hallucination rate across ten major models tested.
What We Don’t Know
AWS’s public launch note does not state Grok 4.3’s pricing or context window directly; those figures come from the model card and from secondary coverage. The launch materials reviewed here do not detail how the Mantle inference engine differs technically from Bedrock’s standard runtime beyond the stated goal of price performance, nor whether other models will migrate to the Mantle path over time. Regional availability beyond the three US regions listed on the model card is not specified.
Analysis
The practical significance of the launch is less about a single new model than about how AWS is choosing to serve it. By routing Grok 4.3 through a dedicated bedrock-mantle endpoint with OpenAI-compatible Chat Completions and Responses APIs — and explicitly not through Bedrock’s long-standing Converse and Invoke interfaces — AWS is offering xAI’s model in a shape that mirrors how developers already call OpenAI and xAI APIs directly. That lowers switching friction for teams that have standardized on the OpenAI SDK, at the cost of a model that sits somewhat apart from the rest of the Bedrock catalog’s unified tooling. Whether Mantle remains a Grok-specific path or becomes a broader serving tier for reasoning models is the open question the launch raises.