GPT-5.6 Sol reaches Bedrock, with region limits and cache controls

Amazon Web Services says GPT-5.6 Sol, Terra, and Luna are now available through Amazon Bedrock and can be called programmatically through the Responses API. AWS lists Sol in US East (N. Virginia) and US East (Ohio), while Terra and Luna are also listed for US West (Oregon). That makes the Bedrock route materially different from treating the GPT-5.6 family as a single globally interchangeable endpoint: an architecture that requires Sol must first fit its workload, residency policy, and failover plan into those two listed regions.

The AWS implementation is aimed at repeated, multi-step agent traffic rather than a simple one-shot completion endpoint. AWS describes an inference layer that keeps requests in a customer-selected region and applies the customer's IAM policies, VPC boundary, and CloudTrail logging. It also offers prompt caching with explicit cache breakpoints. According to AWS, cached input receives a 90 percent discount and is retained for reuse for at least 30 minutes, which could matter when an agent repeatedly sends stable system instructions, tool definitions, or reference material alongside changing task state.

The product split is worth separating from the provider's performance claims. OpenAI describes Sol as the flagship tier, Terra as a lower-cost tier positioned against GPT-5.5, and Luna as the fastest, most affordable option. OpenAI lists API prices of $5 input and $30 output per million tokens for Sol, compared with $2.50 and $15 for Terra and $1 and $6 for Luna. OpenAI also publishes benchmark and estimated cost comparisons with prior and competing models, but its own notes say the latency and cost estimates are simulations and that real-world results can vary substantially. Those figures are useful selection signals, not a substitute for a controlled workload-specific evaluation.

For teams already standardized on AWS, the new path may simplify identity, audit, and regional-control requirements while retaining access to the GPT-5.6 family. It does not remove operational tradeoffs. AWS says classifier-flagged traffic can be retained for automated abuse detection for up to 30 days, and Sol's current regional availability is narrower than that of Terra and Luna. OpenAI separately describes GPT-5.6 features in its own API, ChatGPT, and Codex surfaces, while AWS specifically documents Bedrock access through the Responses API. The sensible next step is to test tool-loop latency, cache hit behavior, data-handling requirements, and model-tier fallback on the actual workload rather than assuming a benchmark result or an AWS control plane guarantees equivalent application behavior.