A New Model Provider Arrives on Bedrock
Amazon Bedrock just widened its roster in a way that matters for enterprise buyers. Grok 4.3, xAI's frontier reasoning model, is now generally available on the platform, making xAI a Bedrock model provider for the first time. For organizations that have standardized on Bedrock as their managed gateway to foundation models, this removes a real friction: teams that wanted to evaluate Grok previously had to reach outside their governed AWS environment to do it. Now the model sits alongside offerings from Anthropic, Meta, and others within the same procurement, security, and billing perimeter.
The headline for cost-conscious buyers is price. On-demand rates come in at $1.25 per million input tokens and $2.50 per million output tokens, with cached input as low as $0.20 per million, which makes Grok 4.3 the cheapest US-lab frontier reasoning model on Bedrock. There is an important asterisk: that input price applies up to 200,000 tokens, and above that threshold the rate doubles. For workloads that lean on very long contexts, the effective economics require modeling actual token distributions rather than trusting the sticker price.
The Capabilities Enterprises Actually Get
Beyond price, the feature set is aimed squarely at agentic and multimodal use. Grok 4.3 offers a 1 million token context window, configurable reasoning levels that let developers trade depth of thinking against latency and cost, native video input, and support for AI agent tool calling. The configurable reasoning is the quietly significant piece. It acknowledges that not every request deserves maximum deliberation, and it hands developers a dial to match the model's effort to the difficulty of the task, which is exactly the kind of control that production systems need to stay economical.
The million-token context and native video input push the model toward workloads that shorter-context, text-only systems handle awkwardly: reasoning over entire codebases, long document sets, or video content in a single pass. Combined with tool-calling support, this positions Grok 4.3 as a candidate for the agentic applications enterprises are increasingly trying to ship, where the model must hold a great deal of context in view and act on external systems rather than merely generate text in isolation.
Mantle Changes How the Model Is Served
The most technically interesting detail is not the model but the plumbing beneath it. Grok 4.3 runs on Mantle, a new inference engine AWS built into Bedrock that speaks the OpenAI API specification rather than Bedrock's standard Converse or InvokeModel protocols. Requests to Grok go to a completely different endpoint, in the form of a dedicated Mantle URL, instead of the familiar Bedrock runtime address. That is a meaningful departure from how Bedrock has historically presented every model behind a single uniform interface.
For engineering teams, this cuts both ways. Speaking the OpenAI specification lowers the migration cost for the large population of applications already written against that API, since much existing code can point at Mantle with minimal change. But it also fractures the tidy abstraction that made Bedrock attractive in the first place, the promise that you could swap models without rewriting integrations. Teams weighing Grok should account for the fact that it does not slot into their existing Bedrock calls as seamlessly as a same-protocol model would.
Why the Compliance Posture Matters
Enterprises do not adopt models on capability alone, and xAI appears to understand this. Grok 4.3 arrives with SOC 2 Type II attestation, HIPAA eligibility, and GDPR compliance for production workloads. For a relatively young lab competing against incumbents with long enterprise track records, leading with a credible compliance story is the price of entry into regulated industries. Healthcare, financial services, and public sector buyers cannot even begin an evaluation without these assurances, and their presence signals that xAI is pursuing serious enterprise deployment rather than developer curiosity.
Delivering the model through Bedrock reinforces that posture. Buyers inherit AWS's own controls around data handling, network isolation, and access management, and they keep their model usage inside the governance framework they have already built and audited. The combination of xAI's certifications and Bedrock's guardrails lowers the barrier meaningfully. It lets a risk-averse enterprise treat Grok as one more governed option in an approved catalog rather than as an exception requiring its own bespoke security review.
The Case for a Multi Model Strategy
We read this launch less as a referendum on Grok specifically and more as further evidence that the sensible enterprise posture is multi model. Adding a competitive, lower-cost reasoning option to Bedrock gives architects room to route workloads by economics and fit rather than defaulting every request to a single expensive model. Cheaper reasoning capacity is useful precisely because so much production traffic does not need the most capable, most costly system, and being able to direct that traffic elsewhere improves the unit economics of AI at scale.
The strategic value is also about leverage and resilience. An organization that can move workloads among several strong models is less exposed to any one provider's pricing changes, capacity limits, or roadmap surprises. Grok 4.3 on Bedrock adds a genuine alternative to that rotation, and its aggressive pricing pressures the whole field. For technology leaders, the practical takeaway is to build abstraction and evaluation harnesses that make model substitution routine, so that each new competitive option like this one becomes a lever to pull rather than a migration to dread.
