GPT-6 Sol and Luna become available through Amazon Bedrock

A new hosting route for the GPT-6 family

AWS has made OpenAI’s GPT-6 Sol and GPT-6 Luna generally available through Amazon Bedrock. Its AWS announcement, posted on 22 September, says the models join GPT-6 Astra in the service. This is separate from the models’ release through OpenAI’s direct API: the relevant news for an AWS customer is the ability to use the models within Bedrock’s access, billing and governance environment.

AWS positions Sol for repeated complex work such as software development and multistep analysis, while Luna is intended for focused tasks at high volume, including extraction, classification and routing. Both support context windows of up to one million tokens, according to AWS. That capacity can be useful, but teams still need to choose the right evidence and watch latency and spend when prompts become very large.

An internal OpenAI factuality evaluation cited by AWS says Sol makes roughly half as many mistakes as GPT-5.6 Sol. The result is a vendor test, not a guarantee for every application. A buyer should compare models on its own repository, documents and decision criteria before changing a production workflow.

Choosing models by task

A customer-service system might use Luna to identify a request type and extract fields, then send exceptional cases to Sol for a deeper investigation. A coding workflow may call Sol when it needs to inspect multiple files, propose a change and run checks. The point is to allocate more capable reasoning where it changes the outcome, rather than paying the same cost for every routine step.

AWS’s accompanying blog describes explicit prompt caching for both models on Bedrock. Reusing stable instructions or a common policy block can help recurring requests, but the actual saving depends on how the application structures prompts and whether the cache is hit. Teams should record cache metrics and task-level cost rather than applying a published token rate to an untested design.

The same model may behave differently within distinct agent frameworks because tool wrappers, prompts and retrieval systems differ. A side-by-side test should hold the workflow constant where possible. Check final accuracy, retry rate, user-visible latency and human correction time. A cheaper call is not a cheaper process if it regularly creates follow-up work.

Controls and data handling on Bedrock

AWS says customers can use its established access controls and audit model invocation activity. The detailed Bedrock post cites IAM, CloudTrail and private network endpoints as relevant infrastructure controls. Those capabilities can simplify governance for teams already operating in AWS, but they must be configured. A model appearing in the Bedrock catalogue does not automatically enforce an organisation’s permissions or logging policy.

The AWS blog says inference data is not used for model training and that customers do not need to opt into sharing data with OpenAI to use these models. It also notes a qualification for automated abuse detection: classifier-flagged traffic may be retained by AWS for up to 30 days, with a zero-data-retention request available through an account team. Data-handling decisions should be based on the current terms for the actual service and account.

Regional availability is another practical boundary. The short announcement directs customers to Bedrock documentation for supported regions, endpoints, inference profiles and pricing rather than promising availability everywhere. Australian organisations should verify routing and data residency before sending regulated or confidential material through a profile that may process outside a preferred location.

A controlled rollout

Start in the Bedrock console or a supported API with a small, representative test set. Confirm model IDs, endpoint support, quotas and feature availability. Run Sol against difficult tasks and Luna against high-volume, narrow ones, then test a routing design that escalates only where needed. Record output quality, latency, cache use and total AWS charges under realistic traffic.

Where an agent can call tools, keep action permissions independent of model choice. A new model may make a better plan, but it should not gain authority to publish, spend money or modify records merely because it is available through a managed service. Require approval for consequential operations and keep an audit trail of tool requests and results.

The Bedrock announcement gives existing AWS customers a new way to deploy two GPT-6 models without moving the application to a separate provider surface. The decision to adopt should turn on workload fit and verified controls, not the model family name alone. The most defensible comparison is a complete task under the same operational constraints and review standards.

Operations teams should also check the limits of observability. CloudTrail can record invocation activity, but the organisation still needs application-level traces that connect a user request to model calls, tool actions and final outcomes without storing more sensitive data than necessary. A failed request should be explainable and safely retried. For a high-volume Luna pipeline, monitor extraction accuracy over time; for a Sol coding agent, retain test results and reviewer sign-off. These measurements help establish whether the new Bedrock models improve real service quality, not merely whether an API call succeeds.