Grok 4.3 on Bedrock adds an OpenAI-compatible, but nonstandard, route

AWS's July 16 developer post says xAI's Grok 4.3 is generally available through Amazon Bedrock. The material is notable less for announcing a new prompt wrapper than for documenting a different serving path: AWS places the model on its Mantle inference engine and exposes OpenAI-compatible Chat Completions and Responses APIs. For teams already standardized on AWS credentials and controls, that creates an additional deployment option for an existing Grok model without making any claim about OutYet release state.

The timing needs care. AWS's Grok 4.3 model card lists a June 15, 2026 launch date, while the detailed implementation post is dated July 16 and identifies the same model ID, xai.grok-4.3. That evidence suggests the newer post is a practical access guide, not evidence of a separate 4.3 revision. It also illustrates why availability claims should be tied to a particular provider surface: the model card calls the lifecycle active, but the usable interface described here is specifically Amazon Bedrock's Mantle endpoint.

The integration is not a drop-in call to the usual Bedrock Runtime APIs. AWS documents a regional base URL of the form bedrock-mantle.{region}.api.aws/openai/v1 and notes that the Responses route is under /openai/v1, unlike the path used by other Responses integrations. The model card lists Chat Completions and Responses support, but not Invoke or Converse. Applications that assume a generic Bedrock client therefore need an OpenAI-SDK-compatible client configuration and endpoint-specific routing rather than only a model-ID substitution.

For agent builders, the documented feature set is broad: a one-million-token context window, text and image input, client-side tool calling, structured outputs, response streaming, and configurable reasoning effort from none through high. The Responses API can return encrypted reasoning content for a later turn, whereas the Chat Completions API does not return reasoning tokens. That makes the Responses path the more natural fit for a stateful, long-running workflow, but the useful effort level remains workload-dependent and should be measured rather than inferred from the provider's positioning.

There are material operating limits. AWS lists in-region availability only in us-west-2, us-east-1, and us-east-2, with no geo or global cross-region option, and shows Standard, Priority, and Flex tiers but not Reserved. Its documentation also cautions that on-demand Mantle throughput scales over time and that quota does not guarantee every request during high demand. Production users should therefore test regional capacity, token cost, and their tool loop before migrating. The sources are vendor documentation, not an independent performance comparison, so they establish the interface and constraints rather than a ranking against competing models.