AWS Expands OpenAI GPT-5.6 on Bedrock to More Than 25 Regions

GPT-5.6 Sol, Terra, and Luna compared

The three models are not interchangeable tiers with different names; AWS positions them for distinct workload profiles.

GPT-5.6 Sol: maximum capability

Sol is the family’s highest-capability option. AWS lists it for autonomous coding, security research, scientific analysis, and deep multi-step reasoning. It is the natural fit when solving a difficult problem matters more than minimizing per-request cost or latency.

GPT-5.6 Terra: balanced production performance

Terra is aimed at general-purpose production applications that need a balance of reasoning ability, performance, and cost. It is the middle option for teams that need more capability than a high-volume fast model but do not want to use the top tier for every request.

GPT-5.6 Luna: fast, economical scale

Luna is designed for latency-sensitive, high-volume workloads such as classification, summarization, and routing. Its positioning makes it better suited to repeated, relatively focused inference than to long-running autonomous or research tasks.

AWS also announced substantial GPT-5.6 pricing reductions effective July 30: Luna’s on-demand inference prices were reduced by 80%, while Terra’s were reduced by 20%.

Shared developer interface and context capacity

The GPT-5.6 family is intended to fit into a common Bedrock integration pattern. The models support text and image inputs, streaming, the OpenAI Responses and Chat Completions APIs, and Bedrock Converse, according to AWS’s cross-Region documentation. Developers using the OpenAI SDK can point their application at Bedrock’s OpenAI-compatible endpoint and use an inference-profile ID for routing.

The family is also associated with long-context workloads. AWS documentation for GPT-5.6 Sol lists a 1-million-token context window, alongside a shorter 272,000-token option. Teams should verify the supported context configuration and current pricing for the specific model and inference profile they plan to use.

How geographic and global profiles route requests

The most important operational distinction is the inference profile prefix.

Geographic profiles

A geographic profile such as us. restricts routing to the geographic boundary defined by that profile. This can provide a clearer data-processing geography, but it also limits the pool of available capacity compared with global routing.

AWS also added India-specific Terra and Luna profiles covering the Mumbai and Hyderabad Regions. For organizations with India-focused residency requirements, those profiles provide a more targeted routing option than a global profile.

Global profiles

A global. profile can route requests to any supported commercial AWS Region according to available capacity. AWS says global routing carries an approximately 10% lower per-token rate than the corresponding geographic option. The trade-off is that it is not appropriate when an application must enforce a fixed processing geography.

In practical terms:

• Choose a geographic profile when residency boundaries are a hard requirement.
• Choose a global profile when the priority is a larger capacity pool, potentially higher throughput, and the stated price advantage.
• Check the destination map and service quotas before treating cross-Region inference as a guarantee of capacity or latency.

Where the announcement fits in the GPT-5.6 rollout

GPT-5.6 Sol, Terra, and Luna became generally available on Amazon Bedrock in July. AWS’s initial positioning covered the same three-part range: Sol for complex reasoning and agentic coding, Terra for balanced production workloads, and Luna for fast, cost-efficient inference.

The cross-Region announcement extends that availability model operationally. It follows other GPT-5.6 updates, including pricing changes and expanded long-context support.

AWS also announced Daybreak access for eligible customers in August. Daybreak Blue provides GPT-5.6 Sol with safeguards calibrated for defensive cybersecurity work, while Daybreak Red provides access to the purpose-trained GPT-5.6 Cyber model. The initial announcement says access requires eligibility and enrollment through OpenAI’s Trusted Access for Cyber program.

Governance and infrastructure context

Bedrock deployments sit within AWS identity and audit controls. AWS IAM governs access, while CloudTrail records actions taken by users, roles, and AWS services for operational and risk auditing.

That governance layer is separate from the routing decision. IAM and logging can help establish who invoked a model and how the workload is managed, while geographic or global inference profiles determine where Bedrock may process the request. Organizations still need to evaluate their own encryption, access policies, logging configuration, and residency requirements.

The announcement also arrived alongside a broader infrastructure story. NVIDIA said on August 20 that its first Vera Rubin racks were powering OpenAI’s training stack. That statement is a contemporaneous company claim rather than an independently audited account, but it is consistent with the previously announced plan for OpenAI to deploy NVIDIA systems for next-generation AI infrastructure.

Bottom line

AWS’s August 20 update makes GPT-5.6 more practical for organizations already building on Bedrock: the three model tiers are available through cross-Region inference in more than 25 Regions, and developers can use familiar OpenAI-compatible interfaces.

The main decision is not simply which model is strongest. Sol is for difficult reasoning and agentic work, Terra for balanced production applications, and Luna for fast, economical volume. The routing choice is equally consequential: geographic profiles favor defined residency boundaries, while global profiles favor capacity flexibility and AWS’s stated per-token discount.