xAI's Grok 4.3 Lands on Amazon Bedrock With 1M-Token Context

Grok 4.3 is now available on Amazon Bedrock, marking xAI's formal debut as a model provider on AWS's managed AI platform. With this launch, xAI joins Amazon Bedrock as a model provider, giving developers more choice as they build generative AI applications across reasoning, agentic, and enterprise workflows. On the surface, it's a straightforward distribution deal. Under the hood, it's a window into a much larger strategic chess match.

What You're Actually Getting

Grok 4.3 is a reasoning-first model with always-on and configurable reasoning effort, offering strong tool use and instruction-following capabilities for building multi-step agents, along with token efficiency for high-volume inference. The model's headline numbers are competitive:

• A 1-million-token context window and support for configurable reasoning efforts: none, low, medium, or high.
• Pricing on Bedrock's on-demand tier comes in at $1.25 per million input tokens and $2.50 per million output tokens, with cached input at $0.20 per million.
• The context window sits at 1 million tokens with a maximum output of 30,000 tokens, which puts it in range for long-document workloads like contract review, case law research, and financial document analysis.
• Beyond text, the model supports native video input and can generate PDFs, spreadsheets, and slide decks , capabilities that push it toward enterprise automation use cases rather than simple chat.

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. Mantle is worth understanding on its own: Project Mantle simplifies and expedites onboarding of new models onto Amazon Bedrock, provides highly performant and reliable serverless inference with sophisticated quality of service controls, unlocks higher default customer quotas with automated capacity management, and provides out-of-the-box compatibility with OpenAI API specifications. That last point matters a lot for migration , you can point your existing OpenAI SDK calls at Bedrock with minimal changes.

The Reasoning Controls Are the Interesting Part

Grok 4.3's reasoning system is more granular than most models currently offer. Reasoning is always active by default. You can configure effort via the reasoning parameter: {"effort": "none"} to disable it, or set it to "low" (default), "medium", or "high". This lets you dial in a cost-vs-accuracy tradeoff per request rather than committing to a single mode.

Reasoning content is encrypted and can be returned by passing include: ["reasoning.encrypted_content"] in the Responses API request. You can send the encrypted content back in subsequent turns to provide reasoning context for multi-turn conversations. This is a practical feature for agent pipelines where you want the model to maintain coherent reasoning chains across steps without leaking raw chain-of-thought tokens to end users.

Here's the minimal setup to get started:

from openai import OpenAI
client = OpenAI()
# Point OPENAI_BASE_URL to: https://bedrock-mantle.us-west-2.api.aws/openai/v1
response = client.responses.create(
    model="xai.grok-4.3",
    reasoning={"effort": "high"},
    include=["reasoning.encrypted_content"],
    input="Summarize this 500-page contract and flag unusual clauses."
)
print(response.output_text)

The Benchmark Story , With Caveats

xAI is leading with Grok 4.3's hallucination numbers, and the claims are grounded in third-party data. Grok 4.3 ranks #1 on the Artificial Analysis Omniscience benchmark for lowest hallucination rate among frontier models, #1 on the Artificial Analysis Tau2 Telecom benchmark for real-world tool-calling performance in customer support scenarios, and #1 on Vals AI Case Law and Corporate Finance benchmarks for complex document understanding.

The Omniscience benchmark specifically measures overconfidence , it penalizes models for confidently giving wrong answers rather than refusing to answer. Grok 4.3 (medium) scores a 16% hallucination rate on AA-Omniscience, and the benchmark rewards correct answers, penalizes hallucinations, and has no penalty for refusing to answer. That's a meaningful edge for enterprise use cases where a wrong confident answer is worse than a "I don't know."

The Bigger Picture: Why This Deal Exists

The Bedrock listing is only part of the story. In February 2026, SpaceX acquired xAI in an all-stock transaction that structured xAI as a wholly owned subsidiary of SpaceX, valuing SpaceX at $1 trillion and xAI at $250 billion, for a combined total of $1.25 trillion. The company behind Grok has been through significant turbulence: Musk announced that xAI would cease to exist as a separate company, with Grok and X now being the AI division of SpaceX.

Analysts at The Register point to a pattern in how AWS structures these deals. AWS isn't primarily trying to sell Grok to enterprise banks , it's trying to sell Trainium to SpaceXAI, a company currently training Grok on something like 550,000 Nvidia GPUs in a Memphis facility. Anthropic committed to more than $100 billion in AWS cloud spending over ten years and up to 5 gigawatts of Amazon's custom Trainium AI chips, with Amazon investing roughly $33 billion total. The Bedrock listing, in this reading, is the visible surface of a deeper infrastructure negotiation.

Who Wins, Who Faces Headwinds

For AWS developers already inside the Bedrock ecosystem, the addition is straightforwardly useful. For AWS shops already invested in Bedrock's security and compliance posture, the ability to call Grok 4.3 without leaving that environment removes a meaningful integration barrier that previously required a separate xAI API key and custom middleware.

The picture for enterprise adoption is more complicated. Of more than 400 vendor-identified AI uses across federal agencies, only three mention xAI or Grok , compared to 234 for OpenAI models and 26 for Anthropic's Claude. Regulated industries that care most about Bedrock's governance features , IAM, PrivateLink, CloudTrail auditing , are also the ones most likely to have concerns about the model's provenance and the organizational instability around it.

There's also a structural irony in the integration itself. Grok 4.3 is available on the openai/v1/responses path on the bedrock-mantle endpoint , a different path from the v1/responses path used by other models. It's a small but telling detail: even within Bedrock's unified interface, Grok sits slightly apart.

What's Now Possible

The combination of Grok 4.3's 1M-token context, low hallucination rate, and Bedrock's enterprise security layer opens up a specific class of workloads that were previously hard to run reliably. Think: contract review, case law research, credit agreement analysis, and financial document Q&A, while delivering consistent quality across conversational AI and multi-turn workflows. These are tasks where the cost of a confident wrong answer is high, and where the ability to process an entire document corpus in a single context window is a genuine capability unlock.

The knowledge cutoff is December 2025, though real-time X platform data access partially offsets that gap for current-events queries. For teams building agents that need to reason over current information alongside long documents, that real-time data access is a differentiator no other model on Bedrock currently offers. Whether that's enough to drive adoption in the face of entrenched Claude and GPT workflows is the open question.