Terra and Luna arrive in both US government regions with 1M-token context; flagship Sol is not part of the launch.
By Ryan Merket · Published
Primary source: AWS
Why it matters #
OpenAI can reach sensitive government workloads through AWS's existing controls, while AWS keeps the infrastructure, billing and agency relationship inside Bedrock.
Amazon Web Services said Monday that OpenAI's GPT-5.6 Terra and Luna models are generally available through Amazon Bedrock in the AWS GovCloud US-East and US-West regions, giving government customers access to the two lower-cost tiers of OpenAI's current model family.
The August 24th release covers Terra, which OpenAI positions as its balanced production model, and Luna, the faster model intended for high-volume inference. GPT-5.6 Sol, OpenAI's flagship tier, is not included in the GovCloud announcement.
AWS says both available models support context windows of up to 1 million tokens, enough to place large codebases, long documents or extended agent histories into a single request. The models also support prompt caching with explicit cache breakpoints. Reused context receives a 90% discount, according to AWS, a material cost control for agents that repeatedly load the same instructions, records or reference material.
The release extends a partnership that put the full GPT-5.6 family on commercial Amazon Bedrock regions on July 24th. The GovCloud expansion arrives one month later, after OpenAI cut Terra's direct API prices by 20% and Luna's by 80% on July 30th. OpenAI currently describes Terra as competitive with GPT-5.5 at half the cost per task and Luna as its lowest-priced model.
OpenAI models behind the government-cloud boundary
AWS GovCloud consists of isolated regions built for US government agencies and organizations handling regulated workloads. AWS says the regions are physically and logically administered by US-citizen personnel and are designed to help customers address requirements associated with FedRAMP High, Department of Defense Impact Levels 4 and 5, criminal-justice data and US export controls.
Availability inside GovCloud matters because agencies and contractors frequently cannot route sensitive prompts, documents or model outputs through ordinary commercial cloud regions. Terra and Luna can now be invoked while processing remains inside either specified GovCloud region, according to AWS.
The models run through Amazon Bedrock's bedrock-mantle
endpoint and the OpenAI Responses API. AWS assigns the model identifiers openai.gpt-5.6-terra
and openai.gpt-5.6-luna
. Developers can use the OpenAI software development kit while pointing its base URL at the Bedrock endpoint, which reduces the integration work required to move an existing OpenAI application onto AWS infrastructure.
AWS documentation lists server-side tool calling, projects and prompt caching among Terra's supported Bedrock features. Those functions put the GovCloud release beyond basic text generation: agencies can use the models as components in coding agents, document-processing systems, structured-data extraction pipelines and applications that call other software tools.
AWS also says prompts and completions sent to GPT-5.6 through Bedrock are not used to train models and are not shared with OpenAI. That separation gives AWS control of the customer relationship, billing and operational layer while adding OpenAI's models to the same service that government customers may already use for models from Amazon, Anthropic, Meta and other providers.
Access still starts in a commercial AWS account
GovCloud customers face an additional activation step. Under AWS's model-access process, a customer must first accept the model's end-user license agreement through the standard AWS account linked to its GovCloud account. The customer can then enable the model from the Bedrock Model Access page inside GovCloud.
The split reflects how AWS handles third-party model agreements across its commercial and government environments. Once enabled, customers can call Terra or Luna through the Bedrock console or Responses API.
AWS is leading with OpenAI's efficiency tiers for this expansion. Terra is aimed at production work where reasoning quality and cost must be balanced, while Luna is built for classification, summarization, routing and other tasks that may generate large request volumes. Sol remains the GPT-5.6 option for the hardest coding, cybersecurity and scientific workloads in supported commercial Bedrock regions.
For AWS, the release broadens Bedrock's appeal to agencies that want OpenAI models without creating a separate deployment and procurement path. For OpenAI, it puts GPT-5.6 into regulated workloads through an established government-cloud vendor, starting with the models most likely to be deployed repeatedly and at scale.