OpenAI's GPT-5.6 Sol, Terra, and Luna Launch on Amazon Bedrock with 272K Context Window

What happened

OpenAI's GPT-5.6 model family — three capability tiers named Sol, Terra, and Luna — is now generally available on Amazon Bedrock, according to a joint post from AWS and OpenAI's Chris Dickens. This marks a new naming convention from OpenAI: the version number (5.6) identifies the generation, while Sol, Terra, and Luna identify durable capability tiers that OpenAI says can advance independently on their own release cadence.

The three tiers

• Sol (openai.gpt-5.6-sol) — the flagship reasoning model, positioned for autonomous coding, security research, scientific analysis, and deep multi-step reasoning. Available in US East (N. Virginia) and US East (Ohio).
• Terra (openai.gpt-5.6-terra) — a balanced model for general-purpose production workloads, trading off reasoning depth against cost. Available in US East (N. Virginia), US East (Ohio), and US West (Oregon).
• Luna (openai.gpt-5.6-luna) — optimized for high-volume, latency-sensitive tasks like classification, summarization, and routing. Same three-region availability as Terra.

All three models share identical technical specs: text and image input, text output, a 272,000-token context window, and support for OpenAI's Responses API. Each model exposes six reasoning-effort levels — none, low, medium, high, xhigh, and max — allowing developers to swap models without changing integration code.

Access and pricing

Models are accessed through a new bedrock-mantle endpoint at https://bedrock-mantle.{region}.api.aws/openai/v1/responses, compatible with existing OpenAI Python and TypeScript SDKs (version 2.45.0 or later for Python). Developers authenticate via AWS IAM, an auto-refreshing token generator, or a short-lived bearer token capped at 12 hours.

According to AWS, pricing matches OpenAI's first-party rates, and usage counts toward existing AWS spending commitments — but no specific per-million-token dollar figures were disclosed in the announcement.

Security and data handling

AWS states that all inference runs inside the customer's VPC under their IAM policies and is logged via AWS CloudTrail, with in-Region processing to support data-residency requirements. Prompts and completions are not used for model training and are not shared with OpenAI by default. Classifier-flagged traffic is retained for up to 30 days for automated abuse detection, with retention behavior configurable by the customer.

Developers can also connect the models to OpenAI's Codex coding agent and use prompt caching to reduce cost on repeated context, according to AWS's documentation.

What this means

This is the first time OpenAI's models have shipped on AWS with a distinct three-tier naming scheme separated from the underlying version number — a structural change that could let OpenAI iterate on Sol, Terra, and Luna independently without forcing a full-version migration for customers. For AWS, it deepens a growing pattern of hosting frontier third-party models (alongside Anthropic's Claude) directly in Bedrock's infrastructure, letting enterprises keep OpenAI-quality inference inside their existing AWS security perimeter and billing.

The lack of disclosed per-token pricing and benchmark scores in this specific announcement is notable — buyers evaluating cost or quality tradeoffs between Sol, Terra, and Luna will need to consult OpenAI's separate pricing pages or run their own comparisons, since the AWS post frames this purely as an infrastructure and API integration story rather than a model-capability disclosure.