xAI Grok 4.3 on Amazon Bedrock (June 2026): Mantle Routing and Rate-Card Math
Quick summary: On June 15, 2026 xAI Grok 4.3 reached GA on Bedrock Mantle at $1.25/$2.50 per 1M tokens — a 10M-in / 1M-out month costs $15 vs $45 on Claude Sonnet 5. Here is the routing math and what breaks on Converse.
Key Takeaways
- On June 15, 2026 xAI Grok 4
- 3 reached GA on Bedrock Mantle at $1
- 25/$2
- 50 per 1M tokens — a 10M-in / 1M-out month costs $15 vs $45 on Claude Sonnet 5
- AWS lifecycle notice (June 30, 2026) — Amazon Bedrock Agents Classic is in maintenance for new customers after July 30, 2026
Table of Contents
AWS lifecycle notice (June 30, 2026) — Amazon Bedrock Agents Classic is in maintenance for new customers after July 30, 2026. Net-new agent builds should use Bedrock AgentCore. Full matrix: lifecycle roundup.
On June 15, 2026, AWS announced xAI Grok 4.3 general availability on Amazon Bedrock, served through the Bedrock Mantle endpoint. Published on-demand list rates: $1.25 / $2.50 per 1M input/output tokens, with $0.20 / 1M cached input (re-verify on Bedrock pricing before you lock a budget).
This is a third frontier option on Mantle beside OpenAI GPT-5.6 and Anthropic Claude — not a drop-in for every Converse-based agent you already run. If you standardize on Bedrock for procurement and IAM, Grok 4.3 is worth a routing slot when long context and aggressive input pricing matter more than staying on a Converse-native path.
What changed on June 15, 2026
Per AWS at GA:
| Attribute | Grok 4.3 on Bedrock |
|---|---|
| Model ID | xai.grok-4.3 |
| API path | Mantle — Chat Completions / Responses (bedrock-mantle.REGION.api.aws) |
| Converse API | Not supported |
| Context | 1M tokens |
| Reasoning | Configurable: none / low / medium / high |
| Regions | us-east-1, us-east-2, us-west-2 |
| Geo / global inference | Not supported at GA |
Implication: teams on bedrock-runtime Converse (Agents Classic, many LangChain paths, Bedrock Flows) cannot swap model IDs silently. Mantle is a separate invoke surface — same billing account, different client configuration and IAM actions.
First-party pricing math (not a client silhouette)
No anonymized engagement is cited for Grok 4.3 adoption. The numbers below are reproducible arithmetic on the published Bedrock rate card — one illustrative monthly shape, labeled as math.
Illustrative month: 10M input tokens + 1M output tokens (agentic production mix with moderate output).
| Model | Input $ | Output $ | Monthly total |
|---|---|---|---|
| Grok 4.3 | 10 × $1.25 = $12.50 | 1 × $2.50 = $2.50 | $15.00 |
| GPT-5.6 Terra | 10 × $2.20 = $22.00 | 1 × $13.20 = $13.20 | $35.20 |
| Claude Sonnet 5 | 10 × $3.00 = $30.00 | 1 × $15.00 = $15.00 | $45.00 |
| GPT-5.6 Luna | 10 × $0.22 = $2.20 | 1 × $1.32 = $1.32 | $3.52 |
Direction: Grok 4.3 is −$30.00 / month (−67%) vs Sonnet 5 and −$20.20 / month (−57%) vs Terra on identical token counts. Luna is still ~4× cheaper on this mix — use Luna for high-QPS classification, not as a Grok substitute for 1M-context reasoning lanes.
Download the full comparison: model-routing-cost-worksheet.csv.
Reproduce this — Open the worksheet CSV and replace the sample 10M/1M columns with your Cost Explorer / CUR input and output totals. Work the Mantle gotchas checklist before promoting a lane. Pair with Bedrock cost optimization: token budgets and model selection.
Opinionated routing: where Grok 4.3 belongs
We recommend: trial Grok 4.3 on new Mantle-native lanes that need long context (large codebases, multi-document RAG, extended chat history) and where list input cost dominates — after a frozen-prompt bakeoff against Terra and Sonnet 5. Keep Converse-native production on Sonnet 5 until Mantle migration is explicit. Keep high-volume thin tasks on Luna (see GPT-5.6 Luna & Terra pricing). Keep hardest agentic / governance-heavy work on Claude Opus 5 where ZDR-default and Guardrails composition matter (Opus 5 on Bedrock).
Trade-off you accept: lower $/MTok and 1M context vs Mantle-only invoke, no Converse tool path, and xAI-specific behavior you must re-validate on your prompt pack.
| Lane | Lean Grok 4.3 | Lean instead |
|---|---|---|
| Long-context doc Q&A, 200K–1M token inputs | Yes | Sonnet 5 if Converse + Guardrails required |
| Everyday mid-tier agents (Mantle-ready) | Trial after bakeoff | Terra (−20% cut still higher $/MTok) |
| High-QPS classification / routing | No | Luna |
| Converse Agents Classic / Flows (no Mantle plan) | No | Sonnet 5 or Nova |
| Regulated ZDR-default flagship agents | Cautious | Opus 5 |
For OpenAI-on-Bedrock Mantle patterns (Responses API, Codex context), see OpenAI models + Codex on Bedrock.
What broke (pattern): Converse assumptions on a Mantle-only model
What broke — A team pointed an existing Bedrock Converse wrapper at
xai.grok-4.3and gotValidationException: model does not support Converse. Root cause: Grok 4.3 is Mantle-only at GA. Detection: CloudWatchModelInvocationErroronbedrock-runtimewith zero Mantle traffic. Fix: migrate invoke tobedrock-mantle.REGION.api.aws/openai/v1, update IAM from runtime-only to Mantle actions, and re-test tool schemas — Chat Completions message format differs from Converse content blocks.
Secondary gotcha: context-tier pricing. List rates ($1.25 / $2.50) may apply only below a context threshold; input cost can roughly double above ~200K tokens on some Bedrock model cards — verify the xAI row before budgeting a 1M-context lane. See mantle-gotchas.md.
Invoke Grok 4.3 on Mantle
Context: OpenAI Python SDK against Bedrock Mantle, region us-east-1, model ID from the Grok 4.3 model card. Auth via Bedrock API key or SigV4 per your path.
# 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
response = client.chat.completions.create(
model='xai.grok-4.3',
messages=[
{'role': 'system', 'content': 'You are a concise support assistant.'},
{'role': 'user', 'content': 'Classify: billing dispute on invoice #4421.'},
],
# reasoning_effort='medium', # none | low | medium | high — verify model card
)
print(response.choices[0].message.content)Regions today: us-east-1, us-east-2, us-west-2. No geo/global inference IDs — capacity-plan inside those Regions. Converse on bedrock-runtime is not an fallback.
What to Do This Week
- Confirm Grok 4.3 is enabled in your Bedrock model access console for the target Region.
- Pull last 30 days of Bedrock token usage; split Converse vs Mantle usage types in CUR.
- Recompute monthly $ with the worksheet CSV at Grok, Terra, Sonnet 5, and Luna rows.
- Pick one long-context lane that is Mantle-ready; freeze 50–100 production prompts; score Grok vs incumbent on pass rate, p95 latency, and $ / completed task at
reasoning_effort=noneandmedium. - Check context-tier pricing if any prompt exceeds ~200K input tokens.
- Update IAM, secrets, and observability dashboards to tag Mantle model IDs — Converse-only Cost Explorer filters will miss Grok spend.
- If you are net-new on agents post–July 30, 2026, route orchestration to AgentCore rather than Agents Classic.
What This Post Doesn’t Cover
- A published FactualMinds latency/quality bakeoff of Grok 4.3 vs Sonnet 5 vs Terra on our internal prompt packs (no harness linked yet).
- Batch, Provisioned Throughput, or Priority tier pricing for Grok 4.3 — confirm on the model card; GA notes reference on-demand only.
- xAI first-party API vs Bedrock residency/compliance comparison beyond pointing at AWS procurement consolidation.
- EU / non-US regional availability beyond the three US Regions AWS listed at GA.
- Guardrails + Knowledge Bases composition on the same Converse call path without a proxy layer.
Use the rate card + your CUR. Promote lane-by-lane.
Related reading
AWS Cloud Architect & AI Expert
AWS-certified cloud architect and AI expert with deep expertise in cloud migrations, cost optimization, and generative AI on AWS.
