# Grok 4.3 on Amazon Bedrock — Mantle gotchas

Use this checklist before routing production traffic to **xAI Grok 4.3** on Bedrock Mantle. Reflects **June 15, 2026** GA — confirm model card and pricing page before committing.

## API surface

| Path | Supported? | Notes |
|------|------------|-------|
| **Bedrock Mantle** — Chat Completions / Responses | Yes | Model ID: `xai.grok-4.3`. Base URL pattern: `https://bedrock-mantle.{region}.api.aws/openai/v1` |
| **Bedrock Converse API** (`bedrock-runtime`) | **No** | Existing Converse-based agents, Flows, and SDK wrappers must be refactored to Mantle or use a different model |
| Geo / global cross-region inference profiles | **No** | Plan capacity inside listed Regions only |

## Regions (at GA)

- `us-east-1` (N. Virginia)
- `us-east-2` (Ohio)
- `us-west-2` (Oregon)

No EU or other Regions at launch — do not assume geo inference IDs work.

## Reasoning effort

Grok 4.3 exposes configurable reasoning: `none`, `low`, `medium`, `high`.

- Higher effort increases latency and often **output token volume** even when input is unchanged.
- Budget by **cost per completed task**, not list $/MTok alone.
- Run a frozen-prompt bakeoff at `none` vs `medium` before promoting agent lanes.

## Context and pricing cliff

- **1M token context window** is advertised — attractive for long-document and multi-turn agent workloads.
- Bedrock pricing may use **context-tier steps** (similar to other long-context models). Published list rates ($1.25 / $2.50 per 1M input/output) may apply only below a threshold; **costs can roughly double above ~200K input tokens** on some model cards.
- **Action:** Re-read [Amazon Bedrock pricing](https://aws.amazon.com/bedrock/pricing/) xAI row and the [Grok 4.3 model card](https://docs.aws.amazon.com/bedrock/latest/userguide/) before budgeting long-context lanes.

## Cached input

- Cache-read input is listed at **$0.20 per 1M tokens** (verify on pricing page).
- Requires Mantle prompt-caching support for sticky prefixes — confirm cache-hit lines appear in Cost Explorer / CUR before assuming savings.

## When NOT to adopt Grok 4.3 yet

- Your stack is Converse-native (Agents Classic, Flows, LangChain Bedrock Converse) and you cannot migrate invoke paths to Mantle this quarter.
- You need Guardrails + Knowledge Bases composition on the same Converse call path without a proxy layer.
- Your workload fits Luna-class volume pricing and does not need 1M context or xAI-specific behavior.
- Legal has not reviewed xAI provider terms for your data classification tier.

## Quick invoke sanity check

```python
# openai>=1.x; OPENAI_BASE_URL=https://bedrock-mantle.us-east-1.api.aws/openai/v1
from openai import OpenAI

client = OpenAI()  # OPENAI_API_KEY + OPENAI_BASE_URL from env / Bedrock API key

response = client.chat.completions.create(
    model='xai.grok-4.3',
    messages=[{'role': 'user', 'content': 'Summarize this ticket in one sentence.'}],
    # reasoning_effort='medium',  # if exposed on your SDK version — verify model card
)

print(response.choices[0].message.content)
```

Swap region in `OPENAI_BASE_URL` to match your residency choice (`us-east-2`, `us-west-2`).