AWS Launches Amazon Bedrock in New Zealand with Claude and Cross-Region Inference

AWS launched Amazon Bedrock in the Asia Pacific (New Zealand) region on March 26, 2026, that is, in Auckland. For local companies, this means that requests to generative models can now be sent from the New Zealand AWS region, and Bedrock will itself distribute inference between Auckland and Australian facilities as needed.

How the launch works

The new AWS region received the identifier ap-southeast-6 and became the starting point for cross-Region inference in Bedrock. In practice, this means that an application can make an API call from Auckland, and the request processing itself will go either to the same region, to Sydney, or to Melbourne. AWS added a separate geographic configuration AU for New Zealand, which keeps processing within the Australia and New Zealand space and gives businesses more bandwidth options.

If a company does not have strict data residency requirements, it can choose a global routing profile. In this mode, Bedrock sends requests to supported commercial AWS regions worldwide to reduce the risk of throttling and handle peak loads.

AWS separately emphasizes that traffic goes over the company's own global network, does not go out to the public internet, and is encrypted when transmitted between regions. Logs of all cross-Region calls remain in the source region, that is, in Auckland.

Which models are available

At launch, the New Zealand region gained access to the Anthropic Claude lineup: Opus 4.5, Opus 4.6, Sonnet 4.5, Sonnet 4.6, and Haiku 4.5. For the global profile, AWS also indicates access to Amazon Nova 2 Lite and a broader set of models from different providers.

Developers do not need to retrain integrations from scratch: profiles work through standard Bedrock APIs, including InvokeModel, InvokeModelWithResponseStream, Converse, and ConverseStream.

• AU geographic profile keeps inference in Auckland, Sydney, and Melbourne
• Global profile provides more available capacity and wider model selection
• Model identifiers change through au. and global. prefixes
• Existing routing for Sydney and Melbourne remains unchanged
• Unified Converse format simplifies switching between models without code rewriting

An important detail: for the Australia-New Zealand profile, Claude models are currently supported, whereas global routing opens a wider set of base models. This separation matters for teams that need to monitor both internal compliance requirements and capacity availability. If predictable processing geography is needed, the AU profile is more logical. If maximum throughput is more critical, the global route with dynamic load distribution is more useful.

Control and limits

Launching in a new region does not eliminate the need for infrastructure configuration. To call models through the AU profile from Auckland, the IAM policy must allow access both to the inference profile in ap-southeast-6 and to the models themselves in all destination regions. AWS recommends a least-privilege schema: allow model invocation in Sydney, Melbourne, and Auckland only when the request goes through the inference profile, not directly.

For companies with Service Control Policies, this is also important because Bedrock must be explicitly allowed to work in all three ANZ regions. There are also operational nuances.

Quotas are calculated at the source region level, so increasing the limit in Auckland applies only to requests sent from ap-southeast-6. AWS measures load on two axes — tokens per minute and requests per minute, and for some Claude models, output tokens are billed at a 5:1 ratio relative to input tokens. For Claude Haiku 4.5 and Nova models, a 1:1 ratio applies.

You can check where exactly a request was processed through CloudTrail by the inferenceRegion field, and monitor latency, errors, and token spending through CloudWatch metrics.

What this means

For AWS, this is not just another region on the list, but an attempt to bring Bedrock closer to enterprise customers in New Zealand who care about both latency and control over processing geography. For the market, this is a good signal: cloud AI platforms are increasingly competing not only on models, but on how carefully they address routing, quota, and compliance issues.