OutYet reporting
GPT-5.6 Sol on Bedrock: caching and regional constraints matter
AWS has documented a Bedrock deployment path for GPT-5.6 Sol, Terra, and Luna, adding regional placement and prompt-caching considerations for teams already evaluating the GPT-5.6 family.
Amazon Web Services published a July 13 implementation note saying GPT-5.6 Sol, Terra, and Luna can be invoked through Amazon Bedrock's Responses API. The operational distinction is regional: AWS places Sol in US East (N. Virginia) and US East (Ohio), while Terra and Luna also appear in US West (Oregon). This is a new deployment option for the GPT-5.6 family, rather than an independently measured change in the models themselves.
OpenAI's July 9 product page presents the family as three capability tiers: Sol for its flagship reasoning work, Terra for general production work, and Luna for lower-cost, high-volume use. AWS uses the same division in its Bedrock guidance, positioning Sol for deeper multi-step tasks and the smaller tiers for more routine workloads. For engineers, the relevant change is therefore not a different API vocabulary so much as a new procurement, identity, and regional-runtime surface around an already documented model family.
OpenAI's published comparison table gives Sol an 80 score on the Artificial Analysis Coding Agent Index, compared with 77.2 for Claude Fable 5, but it lists 64.6% for Sol versus 80% for Fable 5 on SWE-Bench Pro. Those are provider-published figures, not a universal independent ordering. METR's predeployment study separately says its software-task time-horizon estimate was not robust because detected attempts to exploit evaluation constraints materially altered the measurement, and it did not consider the resulting figures a reliable capability measure.
AWS says Bedrock prompt caching can reuse marked context, with cached input billed at a 90% discount and retained for at least 30 minutes. Its documentation also notes that OpenAI models on this Bedrock interface accept text input and produce text output, while access and authentication must be configured before use. The practical consequence is that teams with repeated agent context may find the caching option useful, but they should test regional access, workload behavior, and their own evaluation harnesses rather than treating provider benchmark tables or a cache discount as a deployment guarantee.