Grok 4.3 from xAI now available in Amazon Bedrock
xAI's reasoning-first Grok 4.3 brings configurable reasoning effort and token efficiency to Bedrock — potentially cutting enterprise inference costs significantly.
View original announcement →Visual Summary
graph TD
A{{Grok 4.3 on Amazon Bedrock}}:::announced
B(Amazon Bedrock):::compute
C([Mantle Inference Engine]):::feature
D([Configurable Reasoning]):::feature
E([Tool Calling & Structured Output]):::feature
F([Response Streaming]):::feature
G((Enterprise Apps)):::external
H((AI Agents)):::external
G ==>|"invokes"| B
H ==>|"orchestrates"| B
B ==>|"routes to"| A
A -->|"served via"| C
A -->|"exposes"| D
A -->|"supports"| E
A -.->|"enables"| F
E -->|"powers"| H
D -.->|"optimizes cost"| G
classDef announced fill:#ff9900,stroke:#ec7211,color:#fff,font-weight:bold
classDef compute fill:#e3f2fd,stroke:#1565c0,color:#1565c0
classDef storage fill:#e8f5e9,stroke:#2e7d32,color:#2e7d32
classDef feature fill:#fff3e0,stroke:#e65100,color:#e65100
classDef external fill:#f5f5f5,stroke:#616161,color:#616161
What's New
xAI's Grok 4.3 is now generally available on Amazon Bedrock, marking xAI's debut as a model provider on the platform. Grok 4.3 is a reasoning-first model with configurable reasoning effort levels, strong tool use, and instruction-following capabilities tailored for enterprise and agentic workflows. It runs on Mantle, a new inference engine in Amazon Bedrock purpose-built for price-performance efficiency.
How It Works
- Configurable reasoning effort: Grok 4.3 exposes four reasoning effort levels — none, low, medium, and high — allowing developers to tune the compute-vs-quality tradeoff per request, reducing unnecessary token spend on simpler tasks.
- Mantle inference engine: Grok 4.3 is served via Mantle, a new Amazon Bedrock inference engine optimized for price-performance, distinct from the standard Bedrock inference stack used by other providers.
- Tool calling and structured output: The model natively supports function/tool calling and structured output formats, enabling reliable integration into multi-step agentic pipelines and data extraction workflows.
- Response streaming: Mantle supports streaming responses, allowing applications to display partial outputs progressively and reduce perceived latency in interactive use cases.
- Multi-turn and conversational support: Grok 4.3 is designed to maintain coherence across multi-turn dialogues, making it suitable for chat interfaces, customer support bots, and iterative research workflows.
- Token efficiency: The model is architected for token-efficient inference, meaning it produces high-quality outputs with fewer tokens consumed, directly lowering cost at high inference volumes.
- Standard Bedrock API surface: Grok 4.3 is accessible through the standard Amazon Bedrock API, meaning existing Bedrock integrations (SDKs, Agents, Guardrails) can incorporate it without significant code changes.
Why It's Important
- Expands provider diversity on Bedrock: xAI joining Bedrock as a model provider adds a differentiated reasoning-first model to the platform's roster, giving enterprises more architectural options beyond Anthropic, Meta, Mistral, and Amazon's own Nova family.
- Reasoning configurability reduces cost waste: The ability to dial reasoning effort per request is practically significant — teams can run low-effort reasoning on routine queries and reserve high-effort reasoning for complex tasks, optimizing the cost-quality curve dynamically.
- Enterprise workflow readiness: Explicit design targeting for customer support, legal research (case law), financial document Q&A, and web development signals production-grade reliability rather than a research preview, lowering adoption risk for enterprise teams.
- Mantle engine introduces new infrastructure: The introduction of Mantle as a named inference engine suggests AWS is building specialized serving infrastructure for different model families, which could yield better performance and pricing for future xAI and similar models.
- Agentic AI use cases strengthened: Strong tool use and instruction-following directly address the reliability gap that has historically limited LLM deployment in autonomous agent pipelines, making Grok 4.3 a credible option for production agents.
How It's Different
- Reasoning effort is user-configurable: Unlike most Bedrock models where reasoning behavior is fixed or implicitly determined by the model, Grok 4.3 exposes explicit effort levels (none/low/medium/high), giving developers fine-grained control not available with models like Claude or Nova.
- Runs on Mantle, not standard Bedrock inference: Grok 4.3 is the first model announced on the Mantle inference engine, which is described as purpose-built for price-performance — differentiating it from models running on Bedrock's general inference infrastructure.
- Reasoning-first architecture: While many Bedrock models are general-purpose or instruction-tuned, Grok 4.3 is explicitly positioned as reasoning-first, making it more comparable to reasoning-specialized models like DeepSeek-R1 or Kimi K2 Thinking than to general chat models.
- Token efficiency as a design goal: The emphasis on token efficiency as a first-class property distinguishes Grok 4.3 from larger, more verbose models, making it more cost-competitive at scale even if raw benchmark scores are similar.
- xAI's distinct training lineage: Grok models are trained by xAI with a different data and RLHF lineage than Anthropic, Meta, or Google models, potentially offering different knowledge coverage, tone, and failure modes — useful for ensemble or fallback strategies.
When to Prefer It
- High-volume enterprise inference at cost: When running thousands of daily inference calls (e.g., document processing pipelines, support ticket triage), Grok 4.3's token efficiency and configurable reasoning make it more cost-effective than heavier reasoning models.
- Agentic workflows requiring reliable tool use: When building multi-step agents that invoke external APIs, databases, or code interpreters, Grok 4.3's strong tool-calling and instruction-following reduce agent failure rates compared to models with weaker function-calling fidelity.
- Legal and financial document Q&A: When the use case involves structured reasoning over dense documents — case law research, contract review, or financial statement analysis — the high-effort reasoning mode provides depth while lower modes handle simpler lookups cheaply.
- Customer support automation: When deploying conversational AI for customer-facing support with multi-turn context, Grok 4.3's multi-turn coherence and instruction-following make it suitable for consistent, policy-adherent responses.
- Applications needing streaming UX: When building interactive chat or search interfaces where perceived latency matters, Mantle's streaming support ensures users see progressive output rather than waiting for full completion.
- Diversifying away from a single model provider: When enterprise risk management or vendor diversification policies require fallback or A/B model strategies, adding Grok 4.3 alongside existing Anthropic or Amazon Nova deployments provides a credible alternative with minimal integration overhead.
- Web development assistance workflows: When powering AI-assisted coding or web development tools where structured output (e.g., JSON, HTML scaffolding) and tool use are central, Grok 4.3's native structured output support simplifies integration.
Availability
- GA status: Grok 4.3 is generally available on Amazon Bedrock as of June 15, 2026 — this is not a preview release.
- Region availability: Specific supported regions are listed in the Amazon Bedrock regional compatibility documentation; the announcement does not enumerate regions explicitly, so users should consult the official region availability page before deployment planning.
- Inference routing options: Like other Bedrock models, Grok 4.3 supports In-Region, Geographic (Geo), and Global cross-region routing options, enabling compliance-aware or throughput-optimized deployment configurations.
- Inference engine: Runs on Mantle, a new Bedrock inference engine; pricing details specific to Mantle/Grok 4.3 are not disclosed in the announcement and should be verified in the AWS pricing console or model detail page.
- API access: Accessible via the standard Amazon Bedrock API and SDKs; the model detail page in Bedrock documentation provides the model ID and request schema.
- EULA requirement: As with other third-party Bedrock models, users must accept xAI's End User License Agreement via the Bedrock Model Access page before invoking the model; GovCloud users must complete EULA acceptance through a linked standard region account first.
