Amazon Bedrock Adds OpenAI GPT-5.6 Model Family
Amazon Bedrock made OpenAI's GPT-5.6 Sol, Terra, and Luna generally available on July 13, 2026, giving AWS customers access through the OpenAI Responses API on the bedrock-mantle endpoint. According to AWS, the three tiers target deep reasoning, balanced production workloads, and latency-sensitive high-volume inference, respectively. AWS lists a 272K-token context window, text and image inputs, configurable reasoning effort, and prompt caching across the family.
Amazon Bedrock has made OpenAI GPT-5.6 Sol, Terra, and Luna generally available. The models are available through the OpenAI Responses API on Bedrock's bedrock-mantle endpoint, according to AWS's July 13 announcement and a detailed AWS Machine Learning Blog post.
AWS describes Sol as the flagship tier for autonomous coding, security research, scientific analysis, and multi-step reasoning. Terra is positioned for general production workloads balancing reasoning performance and cost, while Luna is intended for high-volume, latency-sensitive uses including classification, summarization, and routing.
Model access and deployment details
AWS lists the following model IDs and regional availability:
- •Sol: openai.gpt-5.6-sol, available in US East (N. Virginia) and US East (Ohio).
- •Terra: openai.gpt-5.6-terra, available in US East (N. Virginia), US East (Ohio), and US West (Oregon).
- •Luna: openai.gpt-5.6-luna, available in the same three regions as Terra.
According to the AWS blog, all three models accept text and image inputs, generate text output, and provide a 272K-token context window. The API also exposes reasoning-effort settings of none, low, medium, high, xhigh, and max, allowing applications to vary inference effort without changing their API integration.
AWS states that pricing matches OpenAI's first-party rates and that usage can count toward existing AWS commitments. The company also reports that prompts and completions sent through this Bedrock integration are not used to train models and are not shared with the model provider.
Caching and security claims
AWS reports that the GPT-5.6 models support prompt caching with explicit cache breakpoints. Under AWS's stated pricing terms, cached repeated context receives a 90% discount. This is particularly relevant to agentic systems that repeatedly submit large system prompts, codebase context, tool outputs, or accumulated task state.
Amazon's product announcement further describes the service as using in-region data processing and hardware-enforced security with zero operator access during inference. These are AWS product claims, and teams with regulated or sensitive workloads will still need to evaluate regional availability, retention settings, identity controls, and their own data-handling obligations.
The release follows AWS's June rollout of GPT-5.5, GPT-5.4, and Codex on Bedrock, according to Amazon's July update. The broader pattern for managed-model platforms is that tiered model families give engineering teams a practical routing choice: reserve the most capable reasoning tier for complex tasks, while directing routine or high-throughput requests to lower-cost inference tiers. In this release, the common Responses API and reasoning controls reduce the integration differences among those choices.
Key Points
- 1Bedrock now offers three GPT-5.6 capability tiers through one Responses API, enabling workload routing without separate model-provider infrastructure.
- 2AWS lists a 272K-token context window and configurable reasoning effort across all tiers, simplifying application-level model substitution.
- 3Prompt caching's reported 90% cached-input discount can materially affect agent workloads that repeatedly send long context and tool traces.
Scoring Rationale
General availability of a frontier OpenAI model family on Amazon Bedrock is consequential for teams standardizing on AWS governance, billing, and regional controls. The shared API, long context window, reasoning controls, and caching economics are directly relevant to production agent and inference-system design.
Sources
Primary source and supporting public references used for this report.
Practice with real Retail & eCommerce data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Retail & eCommerce problems
