Grok 4.3 lands on Bedrock, xAI finally squeezes onto AWS’s model shelf

Today (June 17), AWS announced that xAI's Grok 4.3 has officially landed on Amazon Bedrock, making xAI an “official Bedrock model provider.” Rumors about this have been circulating since May this year, when Business Insider quoted insiders saying AWS was negotiating with xAI—didn’t expect it to be finalized in just over a month.

If you only look at the press release, it might feel like just another routine “model joins a cloud platform” move. But in the context of 2026, it’s not that simple—Bedrock’s model shelf already has Anthropic, Meta, Mistral, Cohere, Amazon Nova, and OpenAI’s flagship models and Codex (added at the end of April). Now Grok is in, too. Bedrock has almost collected all the top closed-source models’ “seven dragon balls,” with only Google’s Gemini still sticking to its own Vertex AI.

What exactly is being launched this time

Grok 4.3 is a small-step iteration after xAI’s Grok 4 series, positioned as “general reasoning + Agent workflow.” Unlike Grok 4.1 Fast released early this year, which focused on low-latency reasoning, 4.3 takes the comprehensive capability route, optimizing context window size, tool call stability, and completion rates for long-chain Agent tasks. On Google Cloud’s Vertex AI, 4.3’s model ID is grok-4.3, while the 4.1 Fast reasoning branch is grok-4.1-fast-reasoning—two lines in parallel.

AWS’s positioning this time is clear—three scenarios:

• Complex reasoning: Tasks such as math, code, or multi-step logic that require “thinking carefully first”
• Agent workflows: Long task scheduling, tool calling, memory management
• Enterprise-level generative applications: Customer service, document processing, knowledge base Q&A—high-volume work

Of note, AWS’s announcement specifically mentions Grok 4.3 can be used with Bedrock AgentCore. AgentCore is AWS’s key Agent development platform this year, providing intelligent memory, gateway acceleration, authorization policies, tool discovery, and observability—a complete suite of support features. With Grok 4.3 now available, enterprise developers have another “brain” option for building Agents.

Why xAI is willing to join, and why AWS accepts

There are several layers of negotiation here.

First, xAI. Musk’s company has long been criticized for “good models but too few distribution channels”—aside from native integration in X (Twitter) and xAI’s own API, its penetration in the enterprise market is very limited. By contrast, Anthropic, with AWS and GCP dual-cloud distribution, has racked up enterprise revenue in the tens of billions this year. Grok wants to tap into the enterprise market, and cannot avoid aggregator platforms like Bedrock. AWS solves not just traffic but compliance—Bedrock comes with Guardrails, SOC2, HIPAA compliance stack, which xAI can't build itself in the short term.

Second, AWS. In last year’s Q3 earnings call, Andy Jassy said he wanted Bedrock to become “the largest inference engine in the world,” aiming for parity with EC2 someday. To get there, you need a complete model shelf. April’s addition of OpenAI was already a landmark move—since AWS has heavily invested in Anthropic, many assumed OpenAI would never join Bedrock. But business is business; once OpenAI models launched on Bedrock (preview version), AWS solidified its “neutral model marketplace” persona. Adding Grok now sends a clear signal: I don’t take sides—if your model sells, I’ll offer it.

There’s another subtle detail—Musk and Bezos don’t have a great personal relationship, with SpaceX and Blue Origin being direct competitors in rockets. But business collaboration outweighing personal grievances shows how urgently xAI needs enterprise channel access.

How it competes with other Bedrock models

This is what developers care about most. Here’s a rough division of Bedrock’s main models:

| Model | Strengths | Weaknesses | |------|-----------|------------| | Claude 4.x | Code, long documents, Agent stability | High price | | GPT-5 series | Comprehensive reasoning, multimodal | Short time on Bedrock, tight quotas | | Llama 4 | Open-source, fine-tunable, cheap | Reasoning limit below closed-source flagships | | Mistral Large 3 | European compliance, cost-performance | Smaller ecosystem | | Amazon Nova | Deep AWS integration | Medium-tier model strength | | Grok 4.3 | Real-time info, reasoning speed, “daring” | Limited enterprise scenario validation |

Grok 4.3’s differentiators lie in two places: its linkage with X platform data—giving real-time social corpus capabilities others can’t match, useful in public sentiment, finance, and marketing applications. And xAI’s emphasis on “low-restriction alignment,” meaning in enterprise internal knowledge bases or security research, Grok’s refusal rate is indeed lower than Claude and GPT—a plus for many B2B users.

But could it steal Claude 4.x’s code-task share on Bedrock? Short-term, I doubt it. Claude has been running on AWS for nearly two years—enterprise prompt engineering, evaluation suites, and CI/CD integrations are all built around Claude, making migration costly. Grok 4.3’s more likely opportunity is incremental—new projects and real-time data analysis scenarios.

Practical info for developers

Joining Bedrock means you can call Grok 4.3 directly with the familiar AWS SDK, IAM authentication, AWS account billing, without separately applying for xAI’s API key. For teams already in the AWS ecosystem, integration cost is near zero.

If you’re not in AWS and don’t want to juggle multiple API keys, OpenAI Hub already supports Grok 4.3, and you can use OpenAI-compatible calls, directly from inside China without proxies:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.openai-hub.com/v1",
    api_key="your-hub-key"
)

resp = client.chat.completions.create(
    model="grok-4.3",
    messages=[
        {"role": "system", "content": "You are a senior backend engineer."},
        {"role": "user", "content": "Write a Go HTTP client wrapper with rate limiting and retries."}
    ],
    temperature=0.3,
    max_tokens=2048
)

print(resp.choices[0].message.content)

If building Agents, it’s better to explicitly set tool_choice. Grok 4.3’s hallucinations in tool selection are much improved over 4.1, but default auto mode can still occasionally skip tools and jump straight to an answer.

Streaming calls use standard SSE:

stream = client.chat.completions.create(
    model="grok-4.3",
    messages=[{"role": "user", "content": "Explain the differences in expert routing implementations in MoE architectures"}],
    stream=True
)

for chunk in stream:
    delta = chunk.choices[0].delta.content
    if delta:
        print(delta, end="", flush=True)

Some unspoken but noteworthy things

First, pricing. AWS’s announcement didn’t give exact pricing for Grok 4.3, but it usually matches xAI’s direct pricing. Bedrock typically offers batch discounts and Provisioned Throughput options. Based on xAI’s current API pricing, 4.3 on Bedrock will likely be about $3 per million input tokens and $15 per million output tokens, slightly lower than Claude Sonnet 4.5.

Second, regional availability. AWS didn’t specify which Regions will launch first, but based on OpenAI models’ rollout in April, likely us-east-1, us-west-2, and eu-west-1 first, with Asia-Pacific later. For China-based teams building overseas products, you’ll need to plan for latency if using AWS to call Grok 4.3.

Third, whether xAI has any exclusivity clause. Anthropic had partial exclusivity with AWS investment, which later eased. For xAI, joining Bedrock seems non-exclusive—on Vertex AI, Grok 4.3 is already available. This is good for developers—means you can price-shop across clouds.

Fourth, the impact on Anthropic. This is the most interesting point. AWS’s heavy investment in Anthropic, yet bringing in OpenAI and xAI, effectively erodes Claude’s moat on Bedrock. Anthropic’s share may decline—but AWS has likely calculated that growing Bedrock’s overall market is better than betting solely on Claude.

In conclusion

Grok 4.3 joining Bedrock isn’t exciting model-wise—4.3 isn’t revolutionary. But in terms of industry landscape, its signaling matters more than its tech: “supermarketization” of closed-source models is now a done deal. There will no longer be “one cloud only sells one model” scenarios. Developers’ core competitiveness is shifting from “choosing the right model” to “combining and dynamically switching models.”

That’s why aggregator API platforms are increasingly popular—whether it’s AWS Bedrock’s giant self-run, or third-party aggregators like OpenAI Hub, the essence is lowering developers’ model-switching cost. When Grok, Claude, GPT, Gemini, and DeepSeek can all be called with the same key and interface, the real question worth time is: whose model best suits the current task?

References

• OpenAI models, Codex land on AWS - iThome: A detailed report from April on AWS expanding cooperation with OpenAI and the first launch of OpenAI models on Bedrock—important background for understanding Grok’s addition.