xAI's Grok 4.3 Arrives on Amazon Bedrock as the Cheapest Major-Lab Model on AWS

When xAI slipped Grok 4.3 onto Amazon Bedrock in mid-June, it did not arrive as a headline-grabbing frontier launch. It arrived as a price. At $1.25 per million input tokens and $2.50 per million output tokens, xAI's mid-tier reasoning model is now the cheapest major-lab model on Amazon's managed AI platform at these specifications, undercutting rivals from Anthropic and OpenAI by a wide margin and landing in enterprise budgets at exactly the moment those budgets are stretched thinnest.

AWS confirmed general availability on June 15, 2026, with xAI joining Bedrock as a model provider for the first time. Two days later, xAI published its own announcement, framing the move as a bet on cost-efficiency rather than raw capability. "Grok 4.3 sits on the Pareto frontier for intelligence versus cost, delivering 2-10x more intelligence per dollar than other frontier models," the company wrote. The pitch is unambiguous: for the growing share of enterprise workloads that do not need top-tier agentic performance, Grok 4.3 offers frontier-adjacent quality at a fraction of the running cost.

The pricing story

The numbers are the news. Grok 4.3 on Bedrock runs $1.25 per million input tokens and $2.50 per million output tokens, with cached input priced at just $0.20 per million. Set that against the competition on the same platform and the gap is stark. Claude Sonnet 5, which Anthropic launched June 30 at introductory pricing, costs $10 per million output tokens, four times Grok's rate. OpenAI's GPT-5.5 runs $15 per million output tokens, six times as much. For budget-constrained teams running high-volume inference, the difference compounds fast.

Notably, Bedrock's rates are dramatically lower than xAI's own direct API, where Grok 4.3 costs $3 per million input and $15 per million output. That inversion, where the reseller undercuts the maker, is unusual and points to how aggressively AWS and xAI are courting enterprise volume through the Bedrock channel.

The model ships with a 1-million-token context window, a maximum output of 30,000 tokens, and configurable reasoning effort that developers can dial from "none" for maximum speed to "high" for the hardest problems. It accepts text and image input and returns text. Under the hood, it runs on Mantle, a new Bedrock inference engine AWS describes as purpose-built for price performance, with support for tool calling, structured output, and response streaming.

xAI leaned on third-party benchmarks to make the quality case. The company says Grok 4.3 ranks first on Artificial Analysis's Omniscience benchmark, claiming the lowest hallucination rate among frontier models, and first on the Tau2 Telecom benchmark measuring real-world tool-calling in customer-support scenarios. It also cites top marks on Vals AI's Case Law and Corporate Finance evaluations, the kind of dense-document work AWS is steering the model toward: contract review, case law research, credit-agreement analysis, and financial document Q&A.

The multi-model Bedrock strategy

Grok's arrival is less about any single model than about what Bedrock is becoming: a single procurement counter for nearly every major AI provider that matters. With xAI on board, enterprises can now reach Anthropic's Claude, Amazon's own Nova family, Meta's Llama, Mistral, AI21, and Grok through one AWS contract, one billing relationship, and one set of security and governance controls.

That consolidation is the quiet strategic play. For a large enterprise, the friction of AI adoption is rarely the model itself; it is the procurement, the security review, the data-residency sign-off, and the vendor management that each new provider drags along. By absorbing xAI into an existing contract, AWS lets a customer add Grok to its stack without opening a new vendor relationship, and lets teams route different workloads to different models based on cost and capability rather than contractual convenience. A cheap model like Grok 4.3 can handle high-volume, lower-stakes traffic while a premium model handles the hard cases, all billed to the same account.

It also sharpens the price war now defining enterprise AI. The second quarter of 2026 saw companies recoil from ballooning agentic-AI bills, with some teams burning annual budgets in weeks. Anthropic's discounted Sonnet 5 launch was one response to that backlash. Grok 4.3 on Bedrock is another, arriving from a different direction: not a cheaper flagship, but a capable mid-tier model priced to win the workloads that never needed a flagship in the first place. When the same platform offers a $2.50 output model beside a $15 one, the pressure on the expensive end becomes structural.

Where Grok 4.3 sits in xAI's lineup

Grok 4.3 is explicitly a middle child, positioned below the still-in-training Grok 5 and above the original Grok. Grok 5 is expected in the third quarter of 2026, still training on Colossus 2, the xAI supercomputer cluster recently expanded to 1.5 gigawatts. Markets remain skeptical of that timeline: Polymarket contracts on a Grok 5 release closed June 30 at roughly 3 percent probability, an accurate read on how much slack xAI's shipping schedules tend to carry.

That gap is precisely why the Bedrock deal matters for xAI. With its flagship still months out and unproven, distribution through AWS gives the company an enterprise foothold now, on the strength of price and reliability rather than a benchmark crown it does not currently hold.

What to watch

The first thing to watch is whether Grok's benchmark claims survive independent enterprise testing, particularly the low-hallucination assertion, in the document-heavy legal and financial workloads AWS is pitching it for. The second is competitive response: with Grok anchoring the cheap end of Bedrock, watch whether Anthropic and OpenAI extend their introductory discounts or let standard pricing snap back, and whether Google and others answer with their own low-cost Bedrock or platform plays. The third is whether the Bedrock relationship deepens, and specifically whether xAI brings Grok 5 to the platform at launch when it ships, or holds its flagship back for its own API. For now, the message to enterprise buyers is simple: the cheapest major-lab model on AWS is no longer a compromise brand, and the floor on frontier-adjacent AI pricing just dropped again.