Claude Opus 5 on AWS: Workstation Guide for Builders & Business

This is the Workstation developer guide to Claude Opus 5 on AWS — Anthropic’s most capable Opus model, now on Amazon Bedrock and Claude Platform on AWS. It covers capabilities, Bedrock vs Claude Platform, model IDs, Invoke/Converse/Mantle code, governance (ZDR), caching, and how to slot Opus 5 into production agent systems without blowing the token budget.

Source material: AWS Machine Learning Blog — Introducing Claude Opus 5 on AWS (Jul 2026). Specs cross-checked with the Bedrock model card and Anthropic pricing pages. Verify current regions and prices in AWS/Anthropic docs before production.

1. What shipped

On 24 July 2026, Claude Opus 5 became available on AWS via:

1. Amazon Bedrock — unified inference with Guardrails, Knowledge Bases, IAM, and zero data retention (ZDR) by default.
2. Claude Platform on AWS — Anthropic-native APIs/console experience, AWS billing and auth; ZDR available on request.

Anthropic positions Opus 5 as a step-change for agentic coding, long-running agents, and document-heavy professional work, with intelligence that (per Anthropic) matches Fable 5 in many domains at Opus-tier pricing.

2. Specs that matter to engineers

Capability	Opus 5 (Bedrock card)
Context window	1M tokens
Max output	128K tokens
Reasoning	Adaptive thinking on by default (can disable; effort capped high when off)
Knowledge cutoff	May 2026
Modalities	Text + image in; text out
Prompt caching	Yes — min 512 tokens/checkpoint; up to 4 checkpoints; TTL 5m or 1h
APIs	Invoke, Converse, Messages (bedrock-runtime / bedrock-mantle)

List pricing (Claude Platform / Anthropic published): about $5 / MTok input and $25 / MTok output, with large savings via prompt cache hits and batch. Bedrock billable rates can differ by tier/region — check the Bedrock pricing page.

3. Business outcomes (plain language)

• Fewer stuck agents — Opus 5 is built to push through long jobs, recover from tool failures, and push back on bad instructions.
• Higher-trust coding assistants — better navigation of large codebases; closer to “senior engineer” behaviour on hard refactors.
• Document ops — compliance packs, financial analysis, multi-doc reports with deeper consistency.
• Data governance — Bedrock ZDR + residency geo IDs reduce the “can we put this in a public chatbot?” debate.

For financial services, workflow automation, and multi-day project agents, Opus 5 is a natural lead model. Still pair it with human gates on money, auth, and production deploys.

4. Developer router — when Opus 5 vs others

Routine tickets / format / classify     → Haiku / Sonnet
Default product agent / RAG Q&A         → Sonnet 5-class
Hard coding + long agents + docs        → Opus 5  ← you are here
Peak-critical / Mythos-only paths       → Fable 5 / Mythos 5 (if licensed)

Workstation rule of thumb: put Opus 5 on the planner and lead-builder agents; keep cheaper models on fan-out workers. See our multi-agent team playbook.

5. Bedrock vs Claude Platform on AWS

Concern	Amazon Bedrock	Claude Platform on AWS
ZDR	Default	On request
AWS Guardrails / KB / AgentCore	Native	Anthropic-native feature set
Auth / billing	IAM + AWS bill	AWS Console + unified billing
Best for	Enterprise AWS estates	Teams already on Anthropic APIs

6. Model IDs and geo routing

On bedrock-runtime:

• global.anthropic.claude-opus-5 — max throughput, no residency constraint
• us.anthropic.claude-opus-5 — US/Canada residency path
• eu.anthropic.claude-opus-5 — EU residency path
• au.anthropic.claude-opus-5 — Australia residency path

On bedrock-mantle Messages API, use anthropic.claude-opus-5 with the Mantle regional endpoint. Pick geo IDs in your IaC — do not hard-code global. for regulated workloads.

Regional availability (announcement): includes US East (N. Virginia), Asia Pacific (Melbourne), Europe (Ireland), Europe (Stockholm), and more — confirm the live list in Bedrock docs.

7. Prerequisites

• AWS account with Bedrock model access enabled for Opus 5
• IAM: bedrock:InvokeModel, bedrock:InvokeModelWithResponseStream (and related create/inference permissions as required)
• Python 3.8+, boto3, optionally anthropic[bedrock]
• AWS CLI configured for the target region

8. Code — InvokeModel

import boto3
import json

bedrock = boto3.client("bedrock-runtime", region_name="us-east-1")

response = bedrock.invoke_model(
    modelId="global.anthropic.claude-opus-5",
    contentType="application/json",
    accept="application/json",
    body=json.dumps({
        "anthropic_version": "bedrock-2023-05-31",
        "max_tokens": 4096,
        "messages": [
            {
                "role": "user",
                "content": (
                    "Design a multi-region AWS architecture in Python that "
                    "supports 100k RPS with graceful degradation."
                ),
            }
        ],
    }),
)

result = json.loads(response["body"].read())
# Content blocks may include thinking + text — print text blocks
for block in result.get("content", []):
    if block.get("type") == "text" or "text" in block:
        print(block.get("text", block))

9. Code — Converse API (recommended multi-model shape)

import boto3

bedrock = boto3.client("bedrock-runtime", region_name="us-east-1")

response = bedrock.converse(
    modelId="global.anthropic.claude-opus-5",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "text": (
                        "Design a multi-region AWS architecture in Python that "
                        "supports 100k RPS with graceful degradation."
                    )
                }
            ],
        }
    ],
    inferenceConfig={"maxTokens": 4096},
)

blocks = response["output"]["message"]["content"]
print("\n".join(b.get("text", "") for b in blocks if "text" in b))

Converse is usually the cleanest path if you already swap models behind one client.

10. Code — Anthropic Messages via Bedrock Mantle

from anthropic import AnthropicBedrockMantle

client = AnthropicBedrockMantle(aws_region="us-east-1")

message = client.messages.create(
    model="anthropic.claude-opus-5",
    max_tokens=4096,
    messages=[
        {
            "role": "user",
            "content": (
                "Design a multi-region AWS architecture in Python that "
                "supports 100k RPS with graceful degradation."
            ),
        }
    ],
)

for block in message.content:
    text = getattr(block, "text", None)
    if text:
        print(text)

Install with pip install 'anthropic[bedrock]'. Mantle uses SigV4 automatically.

11. Tools mid-conversation

Opus 5 on Bedrock supports adding/removing tools mid-conversation via tool_addition / tool_removal content blocks on role: "system" messages, instead of resending the full top-level tools array. That matters for long agents that unlock tools only after a risk gate. See current Bedrock docs for exact schema.

12. Cyber / safety note

AWS notes Opus 5 improves cyber-related capabilities vs Opus 4.8. In higher-risk areas the service may fall back to Opus 4.8 and notify you; configure fallbacks deliberately in the API so production behaviour is predictable.

13. Production checklist (Workstation)

1. Enable model access; pin geo model ID for residency.
2. Attach Bedrock Guardrails (PII, topics, grounding where used).
3. Turn on prompt caching for large system prompts / tool schemas.
4. Cap maxTokens; log input/output tokens per agent role.
5. Route: Opus 5 for hard paths; cheaper models for volume.
6. Human-in-the-loop for write tools, deploys, and spend.
7. Eval set (golden tasks) before flipping the default model in prod.
8. Wire Review Bot + CI — Opus 5 still needs gates (see wise-developer workflow).
9. Budget alarms on Bedrock spend; prefer cache + batch for bulk jobs.
10. Document the choice in an ADR: “Default lead model = Opus 5 on Bedrock.”

14. How this fits Workstation multi-agent stacks

Brief → Planner (Opus 5) → Builders (Sonnet/Opus mix)
      → Review Bot + Reviewer agent (Opus 5 on hard diffs)
      → Ops / GitOps promote with human gate

Opus 5’s long-running behaviour is a strong fit for Planner and Reviewer agents that must hold Spec + large diffs in context. Do not burn Opus 5 tokens on every lint fix.

15. Getting started tomorrow morning

1. Open Bedrock Playground → select Claude Opus 5 → run one hard coding prompt.
2. Copy the Converse snippet above into a scratch Lambda or notebook.
3. Compare quality vs your current Sonnet/Opus 4.x default on 10 golden tasks.
4. If it wins on hard tasks, flip only the lead-agent model ID in config — not the whole fleet.

Optional: use Bedrock Advanced Prompt Optimization to rewrite prompts against your eval criteria.

16. Closing

Claude Opus 5 on AWS gives enterprises a governed path to frontier Opus-tier coding and agent performance. Bedrock’s ZDR-by-default story matters as much as the model card. Developers should adopt Converse or Mantle, pin geo IDs, cache aggressively, and keep Opus 5 on the critical path — not on every chat. That is how Workstation recommends shipping Opus 5: high leverage, controlled cost, production gates intact.

Published by Workstation.