AWS adds GPT-6 Sol and Luna to Amazon Bedrock

OpenAI’s GPT-6 Sol and GPT-6 Luna are now generally available on Amazon Bedrock, giving customers two models that AWS says cost less than their GPT-5.6 predecessors.

The launch expands AWS’s OpenAI model lineup after its earlier GPT-6 Astra rollout on Amazon Bedrock.

Bedrock customers can use Sol for recurring complex work such as coding, debugging, data analysis and multistep tool use. Luna is geared to focused, high-volume jobs such as extraction, summarization, classification, routing and question answering. The OpenAI model catalog describes Sol as a model for complex coding and agentic workflows and Luna as its most efficient model for focused, high-volume tasks.

Both models have 1.05 million-token context windows, maximum outputs of 128,000 tokens and six reasoning-effort settings: none, low, medium, high, xhigh and max. Medium is the default. They are available through Bedrock Runtime using United States or global cross-Region inference profiles, and through the Mantle endpoint in AWS’s us-east-1 region, according to the OpenAI Bedrock guide.

AWS says both models support explicit prompt caching, which lets an application mark repeated instructions, tool definitions or reference material for reuse instead of processing that content again on every request. The AWS GPT-6 Sol model card and AWS GPT-6 Luna model card list both implicit and explicit caching through the Responses API.

Under global cross-Region Standard pricing for requests with no more than 272,000 input tokens, AWS lists Sol at $2 per million input tokens, $2.50 for cache writes, $0.20 for cache reads and $10 for output. Luna is listed at $0.10, $0.125, $0.01 and $0.50, respectively. The models support only Bedrock’s Standard service tier with pay-per-token billing and no commitment, the AWS model cards say.

AWS says Mantle in-Region and United States geographic cross-Region inference cost 10% more than OpenAI’s first-party Standard base rates, while global cross-Region inference uses the base rates. For requests above 272,000 input tokens, the higher rates apply to the full request: input and cache prices double and output prices rise 50% for both models. OpenAI published pricing shows the same short- and long-context base rates.

The performance and reliability claims have not been independently reproduced. AWS says an internal OpenAI evaluation found Sol made about half as many factual mistakes as GPT-5.6 Sol. AWS says OpenAI’s evaluations found Luna improved factual reliability and communicated results more clearly, but the launch announcement provides neither a numerical Luna result nor an evaluation methodology. AWS also calls the Bedrock inference engine high-performance, secure and reliable at scale without publishing launch-specific uptime, latency or throughput measurements.

On governance, AWS says customers can control model access with IAM policies, audit invocations through CloudTrail and use PrivateLink-powered VPC endpoints to help keep traffic within network boundaries. The company says inference runs on hardware-isolated infrastructure with zero operator access, customer inference data is not used for model training and using the models does not require opting into data sharing with OpenAI. AWS says classifier-flagged traffic used for automated abuse detection can be retained for up to 30 days and processed programmatically; customers can request zero data retention through their AWS account team.

The models can be accessed from the Amazon Bedrock console or supported Bedrock APIs. AWS directs customers to its living documentation for the current regions, endpoints, inference profiles and pricing.