OpenAI GPT-5.6 Sol, Terra, and Luna now generally available on Amazon Bedrock
Three-tier GPT-5.6 family lands on Bedrock with 90% prompt caching discounts and AWS-native billing — Sol leads coding and cybersecurity benchmarks.
View original announcement →Visual Summary
graph TD
A{{GPT-5.6 Sol Terra Luna}}:::announced
B(Amazon Bedrock):::compute
C([Responses API]):::feature
D([Prompt Caching]):::feature
E([Max Reasoning Effort]):::feature
F((Developers)):::external
G(Bedrock-Mantle Endpoint):::compute
H((Autonomous Agents)):::external
I([Three Capability Tiers]):::feature
F ==>|"invokes"| G
G ==>|"routes to"| A
A -->|"hosted on"| B
A -->|"accessed via"| C
A -->|"90% discount"| D
A -->|"Sol configurable"| E
A -->|"Sol/Terra/Luna"| I
H -.->|"coding/cyber/genomics"| F
classDef announced fill:#ff9900,stroke:#ec7211,color:#fff,font-weight:bold
classDef compute fill:#e3f2fd,stroke:#1565c0,color:#1565c0
classDef storage fill:#e8f5e9,stroke:#2e7d32,color:#2e7d32
classDef feature fill:#fff3e0,stroke:#e65100,color:#e65100
classDef external fill:#f5f5f5,stroke:#616161,color:#616161
What's New
OpenAI's GPT-5.6 model family — Sol, Terra, and Luna — is now generally available on Amazon Bedrock, offering three distinct capability tiers from flagship reasoning to cost-efficient inference. All three models are accessible through the Responses API on Bedrock's next-generation inference engine, designed for high-performance, secure, and reliable workloads. Pricing matches OpenAI's first-party rates and counts toward existing AWS commitments.
How It Works
- GPT-5.6 Sol is the flagship reasoning model, optimized for deep multi-step reasoning with a configurable "max reasoning effort" dial that trades compute for accuracy on the hardest problems.
- GPT-5.6 Terra is the balanced tier, delivering performance competitive with GPT-5.5 at approximately half the cost, targeting everyday production workloads at scale.
- GPT-5.6 Luna is the fast, low-cost inference tier designed for high-volume tasks where speed and price efficiency are the primary constraints.
- All three models are accessed via the Responses API on the
bedrock-mantleendpoint, a next-generation inference interface distinct from the standard Bedrock InvokeModel API. - Prompt caching with explicit cache breakpoints allows developers to mark stable context segments; repeated cached tokens are billed at a 90% discount, preventing cost compounding in long agentic workflows.
- The naming convention introduced with GPT-5.6 separates generation (the number) from capability tier (Sol/Terra/Luna), meaning each tier can evolve independently on its own release cadence.
- Bedrock's inference engine provides enterprise-grade security, data residency controls, and reliability SLAs on top of the underlying OpenAI models.
Why It's Important
- Frontier reasoning on AWS infrastructure means enterprises can run OpenAI's most capable models without leaving the AWS security and compliance perimeter, critical for regulated industries handling sensitive data.
- 90% prompt caching discount directly addresses the cost compounding problem in agentic workflows, where the same system prompt or context is repeatedly sent across hundreds of sequential steps.
- Three-tier model family lets teams right-size intelligence to the task — avoiding the cost of using a flagship model for simple classification or routing tasks.
- Usage counts toward AWS commitments (EDPs and similar agreements), allowing organizations to consolidate AI spend under existing AWS contracts rather than managing a separate OpenAI billing relationship.
- Sol's benchmark leadership — 80 points on the Artificial Analysis Coding Agent Index, 73.5% on ExploitBench, and 53.6 on Agents' Last Exam — signals a meaningful capability step for autonomous coding, cybersecurity, and scientific research use cases.
- Long-horizon scientific workloads such as genomics analysis and drug discovery become more tractable with a model that maintains coherence and reasoning quality across extended multi-step pipelines.
How It's Different
- Sol vs. GPT-5.5: Sol scores 73.5% on ExploitBench versus 47.9% for GPT-5.5 at a comparable output-token budget, while using less than half the output tokens and costing roughly one-third less on coding benchmarks.
- Terra vs. GPT-5.5: Terra delivers GPT-5.5-level performance at half the cost, making it the practical replacement for GPT-5.5 in production workloads without a quality regression.
- Explicit cache breakpoints vs. automatic caching: Unlike implicit caching schemes, GPT-5.6's explicit cache breakpoints give developers precise control over which context segments are cached, improving cache hit rates in structured agentic pipelines.
- bedrock-mantle endpoint vs. standard Bedrock: GPT-5.6 uses a dedicated next-generation inference endpoint (
bedrock-mantle), separate from the standardbedrock-runtime, reflecting a new inference architecture optimized for these models. - Durable capability tiers vs. monolithic versioning: The Sol/Terra/Luna naming decouples capability tier from generation, so a future GPT-5.7 Luna can improve without disrupting integrations built against the Terra or Sol tier contracts.
- AWS-native billing vs. direct OpenAI API: Pricing parity with OpenAI first-party rates is maintained while usage consolidates into AWS billing, eliminating dual-vendor invoicing overhead.
When to Prefer It
- Use Sol when building autonomous coding agents that must ship production-quality code, perform multi-file refactoring, or operate over long agentic loops requiring deep reasoning at each step.
- Use Sol for cybersecurity research tasks such as vulnerability discovery, exploit analysis, or red-team automation where ExploitBench-class reasoning is required.
- Use Sol for long-horizon scientific workflows — genomics pipelines, drug discovery, or multi-step biological analysis — where maintaining reasoning coherence across hundreds of steps is critical.
- Use Terra for general-purpose production workloads such as code generation, structured data extraction, document summarization, and content workflows where GPT-5.5-level quality is sufficient and cost efficiency matters.
- Use Luna for high-volume, latency-sensitive tasks such as real-time classification, intent routing, summarization at scale, or any inference pipeline where throughput and per-token cost are the dominant constraints.
- Prefer GPT-5.6 over alternatives when your organization needs to keep AI workloads within AWS's security boundary, leverage existing AWS enterprise discount agreements, or enforce data residency policies through Bedrock's In-Region or Geographic routing options.
- Use prompt caching (any tier) when building agentic workflows with a stable, repeated system prompt or large shared context, to capture the 90% token discount and keep costs linear as agent step counts grow.
Availability
- Status: Generally available (GA) as of July 13, 2026.
- GPT-5.6 Sol regions: US East (N. Virginia) and US East (Ohio) only.
- GPT-5.6 Terra and Luna regions: US East (N. Virginia), US East (Ohio), and US West (Oregon).
- Access method: Responses API via the
bedrock-mantleendpoint; also accessible through the Amazon Bedrock Console. - Pricing: Matches OpenAI first-party rates; usage counts toward AWS commitments (e.g., Enterprise Discount Program).
- Prompt caching discount: Cached tokens billed at 90% discount with explicit cache breakpoints defined by the developer.
- Limitation: Sol has more restricted regional availability than Terra and Luna; no EU or APAC regions are listed at launch.
- Model enablement: Standard Bedrock model access request process applies; GovCloud customers must initiate EULA acceptance through a linked standard AWS account before enabling in GovCloud.
Related Resources
- https://docs.aws.amazon.com/bedrock/latest/userguide/models-region-compatibility.html
- https://us-east-1.console.aws.amazon.com/bedrock-mantle/home?region=us-east-1#
- https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards-openai.html
- https://aws.amazon.com/blogs/machine-learning/openai-gpt-5-6-sol-terra-and-luna-are-now-generally-available-on-amazon-bedrock/
