xAI's Grok 4.3 lands on AWS Bedrock: a reasoning-first model at $1.25 per million input tokens

Frontier models are becoming menu items you pick from a cloud marketplace. On June 15, 2026, AWS added xAI's Grok 4.3 to Amazon Bedrock. Grok 4.3 is a reasoning-first model that keeps reasoning always on, priced at $1.25 per million input tokens and $2.50 per million output tokens. It supports a 1M context and tool calling, tuned for agent workloads. ASAP summarizes how this changes model competition, from the primary announcement.

Three numbers worth watching in the announcement

The core of this launch is compressed into three numbers. First, the listing date of June 15, 2026, the moment xAI stepped outside its own API and onto the AWS distribution channel. Second, the unit price of $1.25 per million input tokens and $2.50 per million output tokens, which sits on the low side for a frontier-class reasoning model. Third, the 1M context, aimed at workloads that load whole long documents at once. Read separately, these are just specs; read together, they reveal which market xAI is targeting. It is a positioning that leads with high-volume unit cost and distribution reach rather than raw benchmark supremacy.

Why they let you turn reasoning down

Grok 4.3 joined Bedrock as a reasoning-first model that keeps reasoning always on. What is notable is that it opens configurable reasoning effort across none, low, medium, and high. Always-on reasoning behaves more consistently across multi-step agent loops, but it also adds tokens and latency. The effort dial hands that trade-off to the user: low for simple classification, high for complex judgment, letting you draw your own cost curve inside a single model. It can be read as a design that turns one model into several tiers.

What the $0.20 cached input tells you

Grok 4.3 is priced on Bedrock on-demand at $1.25 per million input tokens and $2.50 per million output tokens. The point worth noting is that cached input is $0.20 per million, making it one of the most affordable frontier reasoning models on Bedrock. A cache price this low signals optimization for workloads that repeatedly feed the same prompt prefix. Agents replay their system prompt and tool definitions on every call, so a cache discount heavily shapes the real bill. This cache price explains the true cost of running at scale better than the headline unit price does.

What a Korean team should actually weigh

Grok 4.3 is a model with a 1 million token context, native video input, and tool calling support. It is presented as suiting enterprise workloads such as customer support, web development, case-law research, and financial document Q&A. From a Korean practitioner's view, the fact that the distribution path is Bedrock matters. A company already on AWS can attach this model within its existing billing and permission structure, with no separate contract or new vendor review. That said, Korean-language performance, local-region availability, and data-handling location cannot be confirmed from this announcement alone. Validating with your own data before adoption remains necessary.

Look at the inference engine, not just the model

Grok 4.3 runs on Mantle, a new inference engine in Amazon Bedrock. Mantle is built for price performance and supports tool calling, structured output, and response streaming. The signal here is that AWS has begun commoditizing the container that holds the model itself. Once models become cloud commodities, the axis of differentiation shifts from the model to where, at what price, and on which engine it is served. xAI being listed alongside the likes of Anthropic, chosen by task and price, is one scene from that same trend. Going forward, choosing a model becomes a judgment that reads distribution and infrastructure together, not just a performance table.

Source: ASAP summary of AWS's addition of xAI Grok 4.3 to Amazon Bedrock (June 15, 2026; reasoning-first with configurable effort, $1.25 input and $2.50 output per million tokens, 1M context, Mantle inference engine).