Claude Opus 5.5 on AWS (September 2026): Bedrock Adoption Guide

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 September 22, 2026, AWS announced Claude Opus 5.5 on AWS — the first Claude 5.5 model, and Anthropic’s most capable Opus. Bedrock is the path with zero data retention (ZDR) by default, regional residency, and Guardrails / Knowledge Bases composition. Claude Platform on AWS is the Anthropic-native console with AWS billing and auth.

If you already run Claude Opus 5 (GA July 24, 2026), treat this as a lane promotion, not a global model-ID swap. The Bedrock model card pins the architecture: 1M-token context, 128K max output, adaptive thinking always on. Anthropic’s model overview adds the price and the breaking changes. This post is the adoption lens.

What changed on September 22, 2026

Per AWS, Opus 5.5 is the enterprise Opus workhorse: long-running coding and knowledge work, with a report-back style that says what it did, what it found, and what it needs next. Anthropic says it completes tasks with fewer tokens than Opus 5, at a lower price per token, and with cheaper cache reads.

EOL is no sooner than September 22, 2027, with a 6-month legacy period, per the model card.

Specs that change architecture

From the Claude Opus 5.5 model card unless noted:

Guardrails, Knowledge Bases, Flows, and Agents are on bedrock-runtime, not on bedrock-mantle. Mantle supports streaming, prompt caching, abuse detection, count tokens, and computer use. Structured outputs are unsupported on both endpoints, per the card.

Opinionated take: default new long-running coding and knowledge-work lanes to Opus 5.5 on Bedrock (bedrock-runtime, ZDR default, explicit maxTokens). Keep high-QPS chat on Sonnet 5, Haiku 4.5, or Nova. Leave Opus 5 in place until the frozen-prompt bakeoff says 5.5 wins on cost per completed task, not on the version number.

Trade-off you accept: fewer tokens and a lower published per-token price versus a handler rewrite, because thinking cannot be switched off and inter-tool progress text moves into thinking blocks.

Prices: cite Anthropic, then confirm Bedrock

Anthropic’s overview (Bedrock listed as a platform) publishes:

The Bedrock model card does not repeat those dollars. It says Opus 5.5 is billed through AWS Marketplace (charges show under the provider, not under Amazon Bedrock) and points at the Bedrock pricing page. On September 23, 2026 that page’s static extract did not include an Opus 5.5 dollar row. Do not re-forecast from the Anthropic list until the Bedrock sheet shows the rate.

Our previously published Opus 5 Bedrock on-demand pin remains $15 / $75 per million input/output tokens in the calculators. That row stays until the pricing page replaces it.

Invoke Opus 5.5 (Converse, explicit maxTokens)

Context: boto3 Bedrock Runtime, us-east-1, global inference profile. The model card’s sample converse() call omits maxTokens. Leaving it unset reserves toward the 128K ceiling and is a common ThrottlingException cause. Set it.

import boto3

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

response = client.converse(
    modelId='global.anthropic.claude-opus-5-5',
    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))

Profile IDs from the model card:

• Geo: us.anthropic.claude-opus-5-5, eu.anthropic.claude-opus-5-5, au.anthropic.claude-opus-5-5, jp.anthropic.claude-opus-5-5
• Global: global.anthropic.claude-opus-5-5
• Mantle Messages base ID: anthropic.claude-opus-5-5 at https://bedrock-mantle.{region}.api.aws/anthropic/v1/messages

US geo keeps data in US and Canada regions. EU, AU, and JP geos stay inside those geographies. Global has no residency constraint. Confirm the live regional matrix before you hardcode a profile — a static fetch of the card did not preserve the region checkmarks.

What broke (pattern, not a cited client) — An Opus 5 wrapper that set thinking off, or that streamed text between tool calls as a progress UI, does not survive the move. Thinking cannot be disabled, and Anthropic says that inter-tool text arrives in thinking blocks whose text is empty at the default display setting. The progress UI goes quiet until the client requests a display value that returns the text. Forced tool choice returns an error instead of a tool call.

Where Opus 5.5 fits

Rollout: 2-week evaluation

1. Pick three lanes — multi-file coding, one job that runs at least 30 minutes, one document-heavy analysis.
2. Freeze prompts — replay production-representative traces. No ad-hoc demos.
3. Score quality pass rate, latency to a completed task, and cost per completed task. Token list price alone will lie if 5.5 spends fewer tokens, or if always-on thinking spends more.
4. Rewrite the three breaking paths — thinking-off flags, forced tool use, and inter-tool progress streaming — in a branch before the bakeoff, or the scores measure your client bugs.
5. Promote lane-by-lane. Block any pull request that flips the production default model ID before scores land.

Pair agent lanes with AgentCore and token budgets.

What to Do This Week

1. Open the Opus 5.5 model card and copy the geo or global profile that matches your residency rule.
2. Smoke-test Converse in a non-prod account with maxTokens set (start at 1024).
3. Grep callers for disabled thinking, forced tool choice, and progress text read from assistant text between tools.
4. Keep Guardrails on bedrock-runtime. Do not assume Mantle applies them.
5. Schedule the three-lane eval. Leave Opus 5 as the production pin until it finishes.

What This Post Doesn’t Cover

• A confirmed Bedrock dollar rate. Anthropic’s $4 / $20 list is cited above; the Bedrock price sheet did not show an Opus 5.5 row in the extract we checked on September 23, 2026.
• A first-party FactualMinds latency or cost benchmark of Opus 5.5 versus Opus 5. We have not run that bakeoff.
• Claude Platform on AWS console steps beyond the Bedrock-first recommendation.
• Full AgentCore packaging for overnight jobs.

Use this as the adoption frame, then validate on your prompt distribution and quota envelope.

Related reading

• Claude Opus 5 on AWS Bedrock (July 2026)
• Claude Sonnet 5 on AWS (June 2026)
• Claude Fable 5 on AWS: Mythos-Class Guide
• GPT-6 Sol and Luna on Amazon Bedrock (September 2026)
• AWS Bedrock Cost Optimization: Token Budgets and Model Selection
• Amazon Bedrock AgentCore: Production-Ready AI Agents