xAI's Grok 4.3 became available on Amazon Bedrock on June 15, 2026 — the first time xAI has offered a model directly through AWS's managed inference marketplace. The headline numbers: a 1-million-token context window, configurable reasoning depth, and pricing that's 40% cheaper on input and 60% cheaper on output than Grok 4.
The pricing breakdown
On Bedrock's on-demand tier, Grok 4.3 comes in at $1.25 per million input tokens and $2.50 per million output tokens. Cached input drops to $0.20 per million. For comparison, Claude Opus 4.8 sits at $15/$75 input/output, and GPT-4.5 remains in a similar range. Grok 4.3's pricing puts it closer to the mid-tier — a genuine differentiator if the reasoning quality holds up under real workloads.
What Grok 4.3 actually does
This is a reasoning-first model with configurable effort levels: none, low, medium, or high. Setting effort to "none" effectively disables chain-of-thought, giving you a fast completion. High effort means deeper reasoning passes — useful for tasks where correctness matters more than latency, like contract review or multi-step data analysis.
The context window of 1 million tokens with a 30,000-token max output is genuinely large. Long-document workloads — full legal agreements, research corpora, extended codebases — become tractable without chunking hacks.
Native video input is included, which separates it from many Bedrock models. It can also generate PDFs, spreadsheets, and slide decks directly, pushing it toward end-to-end document automation rather than just text generation.
The Mantle engine
Grok 4.3 runs on Mantle, a new inference engine inside Amazon Bedrock tuned for price-performance on reasoning workloads. It supports tool calling, structured JSON output, and response streaming — the standard enterprise integration requirements. AWS's announcement positions Mantle as a general infrastructure improvement, not specific to xAI models.
Where this fits
Bedrock's value is managed infrastructure: no GPU provisioning, pay-per-token billing, and IAM integration for enterprise access control. Putting Grok 4.3 there lowers the friction for organizations already running workloads on AWS who want to experiment with a second reasoning model without standing up a separate API integration.
The model joins Anthropic's Claude family, Meta's Llama, and Mistral on Bedrock — giving AWS customers a wider comparison surface than they had six months ago. The competitive pressure has been good for pricing across the board, and Grok 4.3's aggressive cost structure is a direct response to that dynamic.
For developers building retrieval-augmented applications, legal tech, or financial document pipelines, Grok 4.3 on Bedrock is worth a benchmark run. The pricing makes the experiment cheap.