GPT-5.6 Sol gets an Amazon Bedrock route, with caching and regional controls

AWS announced on July 13 that GPT-5.6 Sol, Terra, and Luna are generally available through Amazon Bedrock. The announcement says pricing matches OpenAI's first-party rates and that usage can count toward existing AWS commitments. For Sol specifically, this is a new procurement and deployment route rather than evidence of a new underlying model revision: the relevant change is that AWS now presents the flagship reasoning tier alongside its Bedrock inference infrastructure.

The operational features described by AWS are more concrete than a generic marketplace listing. AWS says customers can keep requests in a selected AWS Region through In-Region inference, a consideration for workloads with data-residency constraints. It also describes prompt caching with explicit cache breakpoints for repeated system instructions, tool definitions, and reference files; cached input is said to receive a 90 percent discount and remain reusable for at least 30 minutes. Those details matter most for multi-step agents whose repeated context would otherwise be reprocessed on every call.

The Bedrock announcement separates the family into Sol for flagship reasoning, Terra for balanced production work, and Luna for lower-cost, higher-volume work. OpenAI's own GPT-5.6 material adds context for why Sol is the premium option: it describes programmatic tool calling for filtering intermediate data, a max reasoning setting, and an ultra configuration that coordinates four agents by default. The two sources therefore describe different layers of the decision: OpenAI describes model and API capabilities, while AWS describes a managed serving environment and its cost and regional controls.

The performance figures attached to Sol should be read as vendor-reported evaluation results, not as an independent comparison conducted for this report. OpenAI reports 73.5% on ExploitBench versus 47.9% for GPT-5.5 at a comparable output-token budget, while AWS repeats that positioning for Bedrock customers. OpenAI also notes that some latency and cost figures are simulated from production behavior and can differ materially in real deployments, and its cybersecurity footnote says tests used reduced safeguards. These caveats make workload-specific evaluation essential before treating benchmark deltas as a production guarantee.

For technical teams already standardized on AWS, the immediate practical question is whether regional placement, commitment accounting, and cacheable prompt structure offset the integration work compared with calling OpenAI directly. A sensible evaluation would measure cache-hit behavior, end-to-end latency, tool-call reliability, and cost on the team's actual agent traces rather than extrapolating from headline benchmarks. What is verified here is AWS's public availability announcement and the capabilities each vendor describes; what remains unverified is how the Bedrock path performs for a particular account, Region, access configuration, and workload.