Most teams do not pick one model for an application. They pick one model and then quietly stretch it across every step, because switching costs context. The AWS Machine Learning Blog post on GPT-6 Sol and GPT-6 Luna argues that the interesting decision is not which model wins, but which step deserves which level of intelligence.
What actually shipped
Both models are now generally available on Amazon Bedrock, per the AWS post, running on an inference engine the post describes as built for performance, security, and reliability at scale. AWS says both arrive at significantly lower API pricing than their GPT-5.6 predecessors.
The family now has three positions. GPT-6 Astra sits at the top for work where the best result matters more than cost. GPT-6 Sol targets recurring demanding work — feature implementation, debugging, refactoring, code review, data analysis, and multistep processes across tools. GPT-6 Luna targets focused, repeatable tasks at high volume: extraction, summarization, classification, and narrow questions across many users.
That is a tiering story, not a benchmark story. If you have been treating model choice as a single global setting, this is the moment to revisit it.
The evidence AWS cites, and what it does not
On an internal factuality evaluation, the post says OpenAI found GPT-6 Sol made roughly half as many factual mistakes as GPT-5.6 Sol. For Luna, the post cites OpenAI evaluations showing improvements in factual reliability and clearer communication of results. Both claims come from vendor evaluations as described in the AWS post; the supplied material does not include the evaluation methodology, dataset, or task mix, so treat the numbers as directional.
One detail worth acting on: you can adjust reasoning effort per request on Luna to trade quality, responsiveness, and cost. That is a runtime knob, not a deployment decision.
Context reuse is the part people skip
A pipeline that routes cheap calls to Luna and hard cases to Sol still pays to resend the same instructions, tool definitions, policies, and reference material on every hop. The AWS post makes this explicit: reprocessing shared context can erase the savings from picking the right model.
Both models support explicit prompt caching on Bedrock, where you mark prompt content for reuse so later requests process mainly the new input. The post names three fits: coding assistants reusing repository instructions, support apps grounded in the same policies, and document processes applying a consistent extraction schema. If your routing logic is solid but your prompts are rebuilt from scratch each call, caching is probably the larger win.
Governance details that affect your design
Access is governed through IAM policies, invocations are auditable through CloudTrail, and VPC endpoints via AWS PrivateLink keep traffic inside your network boundaries. Inference runs on hardware-isolated infrastructure with zero-operator access, per the post. Inference data is not used for model training, and using these models does not require opting into sharing data with OpenAI.
On retention: classifier-flagged traffic for automated abuse detection is retained by AWS for up to 30 days and processed programmatically, and zero data retention can be requested through your AWS account team. The post does not specify which regions, endpoints, or inference profiles are supported — it points to the Bedrock documentation for that.
A practical way to start
Take one workflow and label each step by consequence. Classification, extraction, and routing go to Luna. Investigation and multistep execution go to Sol. Reserve Astra for the decisions where extra reasoning depth changes the outcome. Then measure total cost to a usable result — output quality, tokens, retries, and latency — rather than price per token.
If you are still deciding how to expose these models, the earlier post on picking the endpoint before the model is the right companion: tiering only pays off once the serving path is settled. The open question is whether your evaluation harness can tell you when a step has outgrown Luna, and the AWS post does not answer that for you.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.