---
title: Claude Opus 5 on AWS (July 2026): Bedrock Adoption Guide for Long-Running Agents
description: On July 24, 2026 AWS shipped Claude Opus 5 on Bedrock with 1M-token context, 128K max output, and ZDR by default. This post covers the Bedrock vs Claude Platform routing call and a lane-by-lane rollout plan.
url: https://www.joinhorizon.co/blog/claude-opus-5-aws-bedrock-2026/
datePublished: 2026-07-30T00:00:00.000Z
dateModified: 2026-10-28T00:00:00.000Z
author: palaniappan-p
category: AI Agents
tags: amazon-bedrock, claude, anthropic, ai-agents, aws-ai, foundation-models
---

# Claude Opus 5 on AWS (July 2026): Bedrock Adoption Guide for Long-Running Agents

> On July 24, 2026 AWS shipped Claude Opus 5 on Bedrock with 1M-token context, 128K max output, and ZDR by default. This post covers the Bedrock vs Claude Platform routing call and a lane-by-lane rollout plan.

> **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](https://www.joinhorizon.co/blog/amazon-bedrock-agentcore-production/). Full matrix: [lifecycle roundup](https://www.joinhorizon.co/blog/aws-service-lifecycle-updates-june-2026/).

> **Update (September 22, 2026):** Claude Opus 5.5 is GA on Bedrock. It is the first Claude 5.5 model, with ZDR by default and adaptive thinking always on. This post stays the July 24 Opus 5 record. Lane-by-lane adoption: [Claude Opus 5.5 on AWS](https://www.joinhorizon.co/blog/claude-opus-5-5-aws-bedrock-2026/).

On **July 24, 2026**, AWS [announced Claude Opus 5 availability on AWS](https://aws.amazon.com/about-aws/whats-new/2026/07/claude-opus-5-aws/), Anthropic’s most advanced Opus model yet, with zero data retention (ZDR) on the Bedrock path. The companion [AWS Machine Learning Blog post](https://aws.amazon.com/blogs/machine-learning/introducing-claude-opus-5-on-aws-anthropics-most-capable-opus-model/) frames it as a step-change for coding, multi-hour agents, and document-heavy enterprise work, with access through **Amazon Bedrock** and **Claude Platform on AWS**.

If you already ship GenAI on Bedrock, treat this as a **routing and evaluation event** and avoid a global model swap. The [Bedrock model card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-opus-5.html) pins the architectural numbers: **1M-token context**, **128K max output**, adaptive thinking on by default, and prompt caching from **512 tokens**. This post is the adoption lens: where Opus 5 fits next to Sonnet 5 and Fable 5, how to invoke it safely, and how to promote lanes without model churn.

## What changed on July 24, 2026

Per AWS and the model card, Claude Opus 5 is positioned for:

- **Coding:** navigating large codebases and adapting strategy mid-task
- **Long-running agents:** hours-to-overnight jobs that recover from errors and keep going
- **Document-heavy professional work:** deeper reasoning over long corpora

Two access paths, same model family:

| Path                       | ZDR               | Best when                                                                                    |
| -------------------------- | ----------------- | -------------------------------------------------------------------------------------------- |
| **Amazon Bedrock**         | **On by default** | You need AWS-native governance, regional residency, Guardrails / Knowledge Bases composition |
| **Claude Platform on AWS** | **On request**    | Teams want Anthropic-native console/APIs with AWS billing and auth                           |

AWS also notes Opus 5 can fall back to **Opus 4.8** in higher-risk cyber areas (configurable via API). Production handlers should treat that as a first-class response path.

For the Sonnet-tier parallel launch, see [Claude Sonnet 5 on AWS](https://www.joinhorizon.co/blog/claude-sonnet-5-aws-bedrock-2026/). For Mythos-class / long-horizon knowledge work, see [Claude Fable 5](https://www.joinhorizon.co/blog/claude-fable-5-aws-bedrock-mythos-class-guide-2026/).

## Specs that change architecture

From the [Claude Opus 5 model card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-opus-5.html):

| Spec             | Value                                           | Why it matters                                                                                 |
| ---------------- | ----------------------------------------------- | ---------------------------------------------------------------------------------------------- |
| Context window   | **1M tokens**                                   | Whole repos + doc packs in one session. You still pay for what you send                       |
| Max output       | **128K tokens**                                 | Long artifacts without chunking. Watch client/stream timeouts                                 |
| Reasoning        | **Adaptive thinking (default on)**              | Extra latency vs “fast” Sonnet calls; effort capped at high if thinking is disabled            |
| Knowledge cutoff | **May 2026**                                    | Newer than prior Opus 4.x cards for recent AWS/Anthropic surface area                          |
| Prompt cache     | **Min 512 tokens**, ≤4 checkpoints, 5m / 1h TTL | Lower threshold than older Opus generations. Rewrite cache points if you were sized for 4,096 |
| Launch           | **July 24, 2026**                               | Active lifecycle; confirm service tiers and regional matrix before hardcoding residency        |

**Opinionated take:** default new **long-horizon coding and overnight agent** lanes to Opus 5 on **Bedrock** (ZDR default + Guardrails). Keep high-QPS chat and cheap classification on Sonnet 5 / Haiku / Nova. Do **not** promote Opus 5 as a drop-in for every Sonnet call. The adaptive-thinking tax and Opus-tier pricing will blow cost caps on short turns.

Trade-off you accept: higher per-task quality and overnight agent reliability vs higher token cost and thinking latency on short prompts.

## Invoke Opus 5 on Bedrock (Converse first)

Prefer the **Converse API** on `bedrock-runtime` with an **explicit `maxTokens`**. Leaving `maxTokens` unset defaults toward the model maximum (**128K**) and silently reserves far more quota than a normal request needs, which is a common cause of unexpected `ThrottlingException`.

Use a cross-region inference profile unless you have a hard in-region requirement:

- Geo: `us.anthropic.claude-opus-5`, `eu.anthropic.claude-opus-5`, `au.anthropic.claude-opus-5`
- Global: `global.anthropic.claude-opus-5`
- Mantle / Messages API base ID: `anthropic.claude-opus-5`

Context: boto3 Bedrock Runtime, region `us-east-1`, inference profile ID.

```python
import boto3

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

response = client.converse(
    modelId='global.anthropic.claude-opus-5',
    messages=[
        {
            'role': 'user',
            'content': [
                {
                    'text': 'Summarize the failure modes in this migration plan and propose a rollback checklist.'
                }
            ],
        }
    ],
    inferenceConfig={'maxTokens': 4096},
)

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

For Anthropic-native Messages ergonomics on Bedrock Mantle (SigV4), AWS documents `AnthropicBedrockMantle` with `model="anthropic.claude-opus-5"`, which is useful when your app already speaks the Anthropic SDK. Guardrails, Knowledge Bases, Flows, and Agents composition stay on the `bedrock-runtime` feature set; Mantle’s supported surface is narrower (see model card).

> **What broke (pattern, not a cited client):** Teams that copy-paste Opus 4.x invoke wrappers into Opus 5 often leave `maxTokens` unset or sized to 64K habits while the card allows **128K**. First production spike: throttling under modest QPS because each request reserved the default ceiling. Fix: set `maxTokens` to the task budget (e.g. 2K to 8K for summaries, higher only for long artifacts) and prefer geo/global inference profiles for headroom.

## Bedrock vs Claude Platform: first-pass routing

| Decision point  | Start with Amazon Bedrock                      | Start with Claude Platform on AWS             |
| --------------- | ---------------------------------------------- | --------------------------------------------- |
| Data retention  | ZDR **default**                                | ZDR **on request**                            |
| Governance      | IAM, residency profiles, AWS-native controls   | Anthropic-native workspace + AWS billing/auth |
| Composition     | Guardrails, Knowledge Bases, Flows, Agents     | Anthropic platform features / console UX      |
| Operating model | Central platform team owns model access in AWS | Product teams prefer Anthropic-style surfaces |

For enterprise control-plane tradeoffs more broadly, see [AWS Bedrock vs OpenAI API for Enterprise](https://www.joinhorizon.co/blog/aws-bedrock-vs-openai-api-enterprise/). For agent runtime choices after Agents Classic maintenance, see [AgentCore vs Amazon Q](https://www.joinhorizon.co/blog/aws-bedrock-agentcore-vs-amazon-q-enterprise-decision-guide-2026/).

## Where Opus 5 fits vs Sonnet 5 and Fable 5

Most portfolios need **lanes**, not one hero model:

| Lane                                                       | Prefer                         | Why                                                                                                                                    |
| ---------------------------------------------------------- | ------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------- |
| Overnight / multi-hour agents, large-repo coding           | **Opus 5**                     | Positioned for long-running recovery and production-quality code adaptation                                                            |
| Agentic tool use at scale with cost sensitivity            | **Sonnet 5** (benchmark first) | Strong agent/coding mid-tier; validate before paying Opus rates                                                                        |
| Mythos-class multi-day knowledge work with safeguard story | **Fable 5**                    | Different retention/refusal contract. Read the [Fable guide](https://www.joinhorizon.co/blog/claude-fable-5-aws-bedrock-mythos-class-guide-2026/) before mixing |
| High-QPS chat / classification                             | **Haiku / Nova**               | Latency and cost; Opus 5 is the wrong default                                                                                          |

Anthropic (via AWS) positions Opus 5 as matching Fable 5’s top-tier intelligence in many domains at **Opus-tier pricing**. Verify that **cost/capability claim on your prompts** before routing everything to Opus.

## Rollout plan: 2-week Opus 5 evaluation

Avoid model churn with a fixed protocol:

1. **Pick three lanes:** (a) multi-file coding, (b) a long-running agent job (≥30 min wall clock), (c) one document-heavy analysis path.
2. **Freeze prompts:** replay production-representative traces; no ad-hoc “try this cool prompt” demos.
3. **Score three metrics per lane:** quality pass rate (human rubric), latency to completed task, **cost per completed task** (not cost per 1K tokens alone).
4. **Instrument fallbacks:** log when cyber/safety paths fall back to Opus 4.8; define accept / retry / escalate.
5. **Promote lane-by-lane:** only move lanes that beat the current baseline on quality **and** stay inside cost/latency budgets.

Pair agent lanes with [Bedrock AgentCore production patterns](https://www.joinhorizon.co/blog/amazon-bedrock-agentcore-production/) and [token-budget / model-selection FinOps](https://www.joinhorizon.co/blog/aws-bedrock-cost-optimization-token-budgets-model-selection/).

{/* TODO: artifact link → factualminds/claude-opus-5-bedrock-eval (prompt harness + CSV scorer) when the eval kit ships */}

## What to do this week

1. Confirm regional / geo / global profile IDs against the live model card for your residency rules.
2. Enable Opus 5 access in a **non-prod** account; smoke-test Converse with `maxTokens: 1024`.
3. Add cache checkpoints only after measuring prompt size ≥ **512** tokens on the sticky prefix.
4. Wire Guardrails on Bedrock if the lane touches customer PII or policy-sensitive content.
5. Schedule the three-lane eval; block any “flip default model ID in prod” PR until scores land.

## What this post doesn't cover

- Official **$/MTok** pricing tables (use [Amazon Bedrock Pricing](https://aws.amazon.com/bedrock/pricing/); rates change).
- A published first-party latency/cost benchmark for Opus 5 vs Opus 4.8 / Sonnet 5 on FactualMinds workloads. The eval harness is not published yet.
- Full AgentCore runtime packaging for overnight jobs (separate guide).
- Claude Platform on AWS console walkthrough beyond the Bedrock-first recommendation.

Use this as an adoption framework, then validate on **your** prompt distribution and quota envelope.

## Related reading

- [Claude Sonnet 5 on AWS (June 2026)](https://www.joinhorizon.co/blog/claude-sonnet-5-aws-bedrock-2026/)
- [Claude Fable 5 on AWS: Mythos-Class Guide](https://www.joinhorizon.co/blog/claude-fable-5-aws-bedrock-mythos-class-guide-2026/)
- [Amazon Bedrock AgentCore: Production-Ready AI Agents](https://www.joinhorizon.co/blog/amazon-bedrock-agentcore-production/)
- [AWS Bedrock Cost Optimization: Token Budgets and Model Selection](https://www.joinhorizon.co/blog/aws-bedrock-cost-optimization-token-budgets-model-selection/)
- [AWS Bedrock vs OpenAI API for Enterprise](https://www.joinhorizon.co/blog/aws-bedrock-vs-openai-api-enterprise/)
- [How to Set Up Amazon Bedrock Guardrails in Production](https://www.joinhorizon.co/blog/how-to-set-up-amazon-bedrock-guardrails-production/)

## FAQ

### Where can I access Claude Opus 5 on AWS?
AWS offers two paths: Amazon Bedrock and Claude Platform on AWS. Bedrock ships Opus 5 with zero data retention (ZDR) enabled by default, regional data residency options, and composition with Guardrails and Knowledge Bases. Claude Platform on AWS gives Anthropic-native platform UX with AWS billing and auth; ZDR is available on request.

### What is the Bedrock model ID for Claude Opus 5?
Base model ID is anthropic.claude-opus-5. For cross-region inference use us.anthropic.claude-opus-5, eu.anthropic.claude-opus-5, or au.anthropic.claude-opus-5 for geo residency, or global.anthropic.claude-opus-5 for maximum throughput with no residency constraint. Prefer Converse with an explicit maxTokens value, because unset maxTokens reserves far more quota than needed.

### Should we move all Opus or Sonnet workloads to Opus 5 immediately?
No. Opus 5 is positioned for coding, multi-hour agents, and document-heavy professional work. High-QPS chat, classification, and cost-capped lanes should stay on Sonnet, Haiku, or Nova until you measure quality, latency, and cost per completed task on your own prompts. Promote lane-by-lane.

### How is Bedrock different from Claude Platform on AWS for Opus 5?
Bedrock is the AWS-managed path: ZDR by default, IAM/PrivateLink-friendly governance, and composition with Guardrails, Knowledge Bases, Flows, and Agents via bedrock-runtime. Claude Platform on AWS mirrors Anthropic’s native console and APIs with unified AWS billing, which is useful when product teams want Anthropic-native workflows and ZDR can be requested separately.

### When should we NOT adopt Claude Opus 5 yet?
Delay if (1) the workload already meets quality targets on Sonnet 5 or Opus 4.x at acceptable cost, (2) you have not fixed maxTokens and prompt-cache checkpoints for the new model’s 512-token cache minimum, (3) latency SLAs cannot absorb adaptive thinking (on by default), (4) cyber/security prompts may hit Opus 4.8 fallback and you have not defined handling, or (5) you need features the model card marks unsupported on your chosen endpoint. Structured outputs were once the common gap; they have been generally available for all Claude models on both endpoints since October 20, 2026.

### What specs matter architecturally for Opus 5?
Per the Bedrock model card (launch July 24, 2026): 1M-token context, 128K max output, adaptive thinking on by default (effort capped at high if thinking is disabled), knowledge cutoff May 2026, prompt caching from 512 tokens with up to 4 checkpoints and 5-minute or 1-hour TTL. Orchestration must handle long jobs, including checkpointing, idempotent retries, and timeouts beyond a single Lambda or API Gateway request.

---

*Source: https://www.joinhorizon.co/blog/claude-opus-5-aws-bedrock-2026/*