OpenAI GPT-6 Sol and GPT-6 Luna are generally available on Amazon Bedrock, joining Astra in the platform’s GPT-6 family. AWS announced Sol and Luna’s availability on September 22, 2026. Teams can access the two models through the Bedrock console or supported Bedrock APIs; AWS’s access guide also describes running Sol, Luna, and Astra through global cross-Region inference from Singapore, Jakarta, Thailand, Malaysia, and Taipei. AWS availability announcement; AWS access guide.
For cloud architects, the deployment question has several parts: which model fits the workload, which API exposes the required features, and how permissions, quotas, regional inference, and monitoring will be configured. An October 9 announcement adds another consideration: Bedrock now supports reasoning summaries for OpenAI models through the Responses API. The available evidence establishes these access paths and capabilities, but does not supply a complete executable configuration. AWS reasoning summaries announcement.
Choosing among Sol, Luna, and Astra
AWS positions Sol and Luna for different kinds of work. Its launch announcement describes Sol as a daily model for recurring complex tasks and software development, while Luna targets focused, high-volume tasks. These are vendor descriptions of intended use, rather than independently measured rankings.
| Model | AWS’s documented positioning | Context information in the supplied evidence |
|---|---|---|
| GPT-6 Sol | Recurring complex tasks and software development, including implementing features, debugging, code review and refactoring, data analysis, and multistep workflows across tools | Up to 1M tokens |
| GPT-6 Luna | Focused, high-volume summarization, extraction, classification, and routing; adjustable reasoning effort to balance quality, speed, and cost per request | Up to 1M tokens |
| GPT-6 Astra | Available on Bedrock alongside Sol and Luna | No context limit supplied |
The AWS launch announcement supplies the Sol and Luna descriptions and context limits. The access guide names Astra as a supported model, but the evidence does not give enough detail to compare its workload fit, pricing, or performance with the other two.
The distinction suggests a useful starting point for evaluation: assess Sol against complex development and tool-based workflows, and Luna against narrower tasks performed at volume. That is an interpretation of AWS’s positioning, not a guarantee that either model will be the best choice for a particular application. Likewise, a stated context limit identifies capacity; it does not establish answer quality across that entire context.
AWS also reports that OpenAI’s internal factuality evaluation found GPT-6 Sol made approximately half as many factual mistakes as GPT-5.6 Sol. The supplied evidence does not include the evaluation methodology, task distribution, or absolute error counts. The result should therefore remain an attributed internal finding, rather than a general claim that error rates will halve in production. AWS launch blog.
Access paths and regional configuration
AWS’s access guide covers both Converse and OpenAI-compatible APIs. Its availability announcement separately confirms console and programmatic access for Sol and Luna. This gives teams documented interface choices, but the supplied evidence does not establish that every feature is available identically through every interface. AWS access guide; AWS availability announcement.
The regional guide specifically names global cross-Region inference from Singapore, Jakarta, Thailand, Malaysia, and Taipei. That wording matters when reviewing deployment requirements: the listed locations describe source access, and the supplied evidence does not enumerate destination Regions or provide a data-residency guarantee.
AWS identifies IAM and SCP permissions, quota sizing, and monitoring inference from the source Region as parts of the setup. A deployment review should therefore cover each of those areas alongside API selection. The evidence does not provide exact policy statements, model identifiers, inference-profile identifiers, quota values, or endpoint configuration, so it cannot support a copy-and-paste deployment recipe. AWS access guide.
For architects, the practical implication is that model availability and operational readiness require separate decisions. The guide establishes an access route; permissions, capacity planning, and monitoring still need concrete configuration for the intended workload.
Reasoning summaries use the Responses API
AWS’s October 9, 2026 announcement documents support for the reasoning.summary parameter through the Responses API. It lets users request a human-readable summary of a model’s reasoning alongside the response, with coding, analysis, and multistep problem solving among the tasks AWS highlights. The summary is returned in the summary array of the reasoning output item. AWS reasoning summaries announcement.
These exact identifiers are useful for integration planning. An application that requests the feature needs to account for the reasoning output item and its summary array alongside the answer. The supplied evidence does not specify accepted parameter values or a complete request payload, so constructing an example would require information beyond the documented material.
AWS says the capability applies to all OpenAI models on Bedrock in all AWS Regions where OpenAI GPT models are available, including in-Region, geographic (GEO) cross-Region, and global cross-Region inference. This is conditional on model availability; it does not establish availability in every AWS Region. The announcement documents a summary capability, without supplying evidence that the summary verifies the correctness of the answer.
Governance, retention, and cost claims
AWS describes several controls around Sol and Luna on Bedrock: model access governed by AWS Identity and Access Management (IAM) policies, invocation auditing through AWS CloudTrail, and virtual private cloud (VPC) endpoints powered by AWS PrivateLink that help keep traffic within network boundaries. It also says inference runs on hardware-isolated infrastructure with zero-operator access. These are AWS’s descriptions of the service, rather than independent security findings. AWS launch blog.
The same source says inference data is not used for model training and that using Sol and Luna does not require opting into data sharing with OpenAI. It also documents a retention qualification: for automated abuse detection, AWS retains classifier-flagged traffic for up to 30 days and processes it programmatically. The training restriction should therefore be read alongside that retention provision, rather than interpreted as a promise of no retention.
On cost, AWS says Sol and Luna have significantly lower API pricing than their GPT-5.6 predecessors and support explicit prompt caching on Bedrock. The supplied evidence includes no prices, quantified savings, or cache configuration details. Those claims identify areas to examine during cost planning, but do not support a numerical estimate of workload savings. AWS launch blog.
Taken together, the announcements establish model access, regional inference options, and a documented mechanism for requesting reasoning summaries. The remaining implementation work is concrete: select the model and interface, configure the permissions and capacity described by AWS, and evaluate the workload against the service’s documented data-handling terms. Vendor positioning and internal evaluation results can inform that assessment, but the supplied evidence does not replace application-specific validation.
