OpenAI GPT-5.6 models now available on Amazon Bedrock

OpenAI GPT-5.6 Sol, Terra, and Luna are now generally available on Amazon Bedrock, enabling developers to access frontier reasoning and inference models through AWS's managed infrastructure without operating separate model systems. The three models serve distinct use cases: Sol handles autonomous coding, security research, scientific analysis, and deep multi-step reasoning; Terra balances reasoning, performance, and cost for general-purpose production workloads; and Luna optimizes for high-volume, latency-sensitive tasks such as classification, summarization, and routing.

All three models support a 272K-token context window, accept text and image input, and output text. They operate with the OpenAI Responses API on the bedrock-mantle endpoint, which developers access via a base URL (https://bedrock-mantle.{region}.api.aws) and the /openai/v1/responses path. The models support reasoning effort levels from none to max, allowing developers to trade latency and cost for result quality on complex, multi-step tasks. Sol and Terra are available in US East (N. Virginia) and US East (Ohio); Luna extends to US West (Oregon). Pricing matches OpenAI's first-party rates, and usage counts toward existing AWS commitments.

Security and data handling follow AWS standards. Every model call runs under AWS Identity and Access Management policies, inside the customer's virtual private cloud, and is logged in AWS CloudTrail. Regional processing keeps requests within the specified AWS Region to support data-residency compliance. For these OpenAI models, classifier-flagged traffic is retained for up to 30 days for automated offline abuse detection; retained inputs and outputs are stored and processed by AWS and are not shared with OpenAI unless the customer opts in. Developers control retention configuration through data retention mode.

To get started, developers authenticate using either an auto-refreshing short-term key derived from AWS credentials (the recommended approach for production) or a short-term key from an environment variable (which expires after at most 12 hours). The OpenAI Python SDK version 2.45.0 or later is required. Developers can call GPT-5.6 models through the BedrockOpenAI client by specifying the model ID (openai.gpt-5.6-sol, openai.gpt-5.6-terra, or openai.gpt-5.6-luna), an input prompt, and optional parameters such as max_output_tokens and reasoning effort. Tool calling is supported, allowing models to request tools and use their results to complete tasks. The new Amazon Bedrock console offers a project-based experience where developers can browse, compare up to three models side by side on capabilities, modalities, context window, pricing, and regional availability, and evaluate models before writing application code.

Prompt caching reduces costs in agentic and multi-step workloads that repeat context between calls. Implicit caching is enabled by default and automatically caches eligible requests without code changes. Explicit caching lets developers mark cache breakpoints to control precisely which parts of a prompt are cached. Both modes allow reuse of processed context—such as system instructions, tool definitions, and reference files—across multiple requests, reducing costs as request volume grows.

OpenAI's release of its GPT-5.6 family on Amazon Bedrock represents a direct integration of frontier models into AWS's managed infrastructure, eliminating the need for developers to operate separate model endpoints. The three-tier approach—Sol for reasoning-heavy tasks, Terra for balanced production work, and Luna for high-throughput, latency-critical inference—allows teams to right-size both capability and cost for different workloads within a single API ecosystem.

The integration preserves AWS's security model: all requests operate within customers' IAM policies, VPCs, and CloudTrail logging, with optional regional residency. This addresses a key concern for enterprise workloads subject to data-governance requirements. Pricing parity with OpenAI's first-party offering, combined with credit toward existing AWS commitments, removes a common barrier to adoption. The prompt caching feature (available in implicit and explicit modes) directly targets the economics of agentic and multi-turn reasoning workloads, which repeat context between calls—a pattern that will likely dominate as developers scale autonomous systems.

No comments yet. Be the first to share your thoughts!