AWS says Amazon Bedrock now supports Claude Opus 5 and Claude Sonnet 5 with inference processed entirely within the Seoul Region, and Claude Sonnet 5 with the same arrangement in Singapore. For cloud architects with local processing requirements, the central change is where inference takes place. The announcement specifies particular models and Regions; it does not establish the same availability for every Claude version. AWS’s regional announcement
Separate AWS documentation describes Claude Opus 5.5’s technical specifications and geographic inference profiles, while an announcement for Claude Fable 5.1 sets out availability and safety-review requirements. Together, these sources show why a Bedrock deployment decision needs to pair the exact model version with its routing boundary, interface and data-retention conditions.
What the Seoul and Singapore rollout covers
The regional announcement identifies the following combinations:
| AWS Region | Models named for in-region inference |
|---|---|
| Seoul | Claude Opus 5 and Claude Sonnet 5 |
| Singapore | Claude Sonnet 5 |
AWS says inference for these combinations is processed entirely within the Region called, allowing users with local processing requirements in South Korea or Singapore to use the models at scale. That is a specific processing-location statement. It should not be extended into a claim about every part of an application’s data handling or its overall compliance.
Version names also matter here. The announcement names Opus 5 and Sonnet 5. The separate Opus 5.5 model card documents a different version and different routing options. Its specifications do not establish that Opus 5.5 is included in the Seoul or Singapore rollout.
Claude Opus 5.5: context, interfaces and lifecycle
AWS’s Opus 5.5 model card lists a September 22, 2026 launch date, a 1M-token context window and a maximum output of 128K tokens. The knowledge cutoff is June 2026.
The context window and output limit are separate specifications: the 1M-token figure should not be read as the maximum response length. Nor does the context capacity itself establish performance on a particular workload. The supplied evidence contains no independent benchmark demonstrating how reliably Opus 5.5 uses that full window.
AWS’s broader Anthropic product page says most new Claude models offer a 1 million token context window by default. The qualification matters: “most” is not a specification for every model. For Opus 5.5, the model card provides the version-specific figure.
The card lists bedrock-runtime and bedrock-mantle endpoints and the Messages, Responses, Chat Completions, Converse and Invoke APIs. AWS recommends bedrock-runtime for new applications whenever possible. The retained specifications identify response streaming, implicit and explicit prompt caching, abuse detection, count tokens and structured outputs for bedrock-runtime.
For explicit prompt caching, the documented limits are a minimum of 512 tokens per cache checkpoint and a maximum of 4 checkpoints per request. Supported time-to-live values are 5 minutes and 1 hour; the fields accepting checkpoints are system, messages and tools. These are configuration constraints, not evidence of a particular latency reduction or cost saving.
The model is offered and billed through AWS Marketplace, with product ID prod-e5xo6z26ie45g. AWS says charges appear under the model provider on the AWS bill and in AWS Cost Explorer. The lifecycle information specifies end of life no sooner than September 22, 2027 and a 6-month legacy period. The earliest permitted end-of-life date should not be treated as an announced retirement date.
Geographic profiles have different boundaries
Opus 5.5’s model card describes these inference profiles:
| Profile identifier | Documented routing boundary |
|---|---|
| <code>us.anthropic.claude-opus-5-5</code> | US and Canada regions |
| <code>eu.anthropic.claude-opus-5-5</code> | EU regions |
| <code>au.anthropic.claude-opus-5-5</code> | Australia regions |
| <code>jp.anthropic.claude-opus-5-5</code> | Japan regions |
| <code>global.anthropic.claude-opus-5-5</code> | Worldwide, with no residency constraints |
The US profile’s inclusion of Canada is a consequential detail for anyone evaluating a requirement limited to the United States. Its us. prefix alone does not convey the full documented boundary.
Likewise, a geographic profile describes a broader boundary than the single-Region processing promised in the Seoul and Singapore announcement. Global routing carries no residency constraint in the model card. Architects should therefore evaluate the selected profile against the required processing geography, rather than infer that geography from the model name or the Region from which a request originates. The evidence does not establish a latency or throughput comparison among these options.
Fable 5.1 adds safety-review conditions
AWS’s Fable 5.1 announcement states that the model is available on Amazon Bedrock and Claude Platform on AWS. On Bedrock, the named options are US Geo CRIS (us.) and Global CRIS (global.) inference profiles. AWS also identifies both bedrock-runtime and bedrock-mantle endpoints in AWS GovCloud (US), and Claude Platform on AWS availability in North America.
Anthropic has designated Fable 5.1 a Covered Model. According to AWS, use requires aws_review mode to be set before invocation. Under that mode, AWS retains prompts and outputs for human safety review within the AWS boundary; sharing the data with the model provider is not required. Keeping data within AWS and retaining it for review are compatible statements. The former does not imply zero retention.
The announcement separately describes Enterprise Frontier Safeguards (EFS), built jointly by AWS and Anthropic. EFS-eligible customers can use Claude Fable 5 and Claude Fable 5.1 with zero data retention on Bedrock and Claude Platform on AWS for internal use through December 31, 2026. Eligibility, internal-use scope and the end date are all material conditions. This qualified arrangement should not be presented as the default for every Fable customer. The supplied evidence does not detail the eligibility criteria or establish what follows that date.
AWS also relays Anthropic’s claim that Fable 5.1 improves on Fable 5 on difficult reasoning tests, including competition mathematics, graduate-level engineering and science questions, and long multi-step problems. No numerical scores, full testing methodology or independent evaluation are supplied, so the claim does not support a quantified performance comparison.
For deployment planning, the practical implication is to assess model version, routing and retention together. The regional rollout supplies a specific inference-location commitment; Opus 5.5’s card supplies version-specific limits and geographic profiles; Fable 5.1’s announcement adds explicit review and retention conditions. None of those facts substitutes for the others when deciding whether a model fits a workload.
