GPT-6 Sol and GPT-6 Luna arrive on Amazon Bedrock

OpenAI's GPT-6 Sol and GPT-6 Luna are now generally available on Amazon Bedrock, AWS said in a blog post, extending the GPT-6 family on the cloud platform beyond GPT-6 Astra, which arrived earlier for the most demanding projects.

Two models, two jobs

According to AWS, the pair covers different points on the intelligence-versus-efficiency curve. GPT-6 Sol targets the demanding work that repeats through a development cycle: implementing features, debugging, refactoring and reviewing code, analyzing data, and carrying multistep processes across tools and applications. GPT-6 Luna is built for focused, repeatable tasks at high volume — extracting information from large document collections, summarizing incoming material, classifying inputs and answering narrow questions for many users or applications.

AWS says both models arrive at significantly lower API pricing than their GPT-5.6 predecessors, which the company frames as the main lever for putting more capable models into regular production use.

Accuracy and the cost of one result

On an internal factuality evaluation, OpenAI found that GPT-6 Sol made approximately half as many factual mistakes as GPT-5.6 Sol. The company also points to clearer reporting about what a model changed, what it verified and what it could not confirm — a detail AWS argues matters for long tasks where teams need to see where human judgment is still required. For GPT-6 Luna, OpenAI's evaluations show improvements in factual reliability, and reasoning effort can be adjusted per request to balance quality, latency and cost.

Mixing models and caching prompts

Because a single application may need different levels of intelligence at different steps, AWS describes routing as a practical pattern: Luna to classify incoming requests, Sol to investigate complex cases and Astra when extra reasoning depth can change a decision. To keep that setup efficient, both new models support explicit prompt caching on Bedrock, letting developers mark reusable content — repository instructions, policy documents, a fixed extraction schema — so later requests process mainly the new input.

Governance and data handling

On the control side, AWS lists access management through IAM policies, auditing of every invocation with CloudTrail, VPC endpoints powered by AWS PrivateLink, and inference on hardware-isolated infrastructure with zero-operator access, meaning AWS operators cannot read prompts or completions during inference. The company says inference data is not used for model training and that using these models does not require opting in to share data with OpenAI; classifier-flagged traffic for automated abuse detection is retained by AWS for up to 30 days, and zero data retention can be requested through an AWS account team. Both models are available in the Bedrock console and through supported Bedrock APIs.