AWS Bedrock GPT-5.6 Sol, Terra, and Luna: Key Differences

Quick answer

Choose Sol when the hardest reasoning or coding quality is the priority, Terra for a balance of capability and cost in everyday production work, and Luna for fast, high-volume tasks where unit cost and latency matter. Then verify the choice on your own evaluation set.

Bedrock tier	AWS positioning	Starting workloads	Model ID
Sol	Flagship reasoning	Complex coding, difficult analysis, high-value agent decisions	openai.gpt-5.6-sol
Terra	Balanced production	General assistants, mixed tool work, extraction that needs judgment	openai.gpt-5.6-terra
Luna	Fast and affordable	Classification, routing, summarization, high-volume structured tasks	openai.gpt-5.6-luna

This table describes product positioning, not a promise that one tier will win every workload. A clear prompt, narrow tool contract, and strong evaluation can matter more than tier size.

What all three share on Bedrock

The current AWS documentation gives Sol, Terra, and Luna a common operational foundation:

• 1 million-token context window;
• text and image input, text output;
• Responses and Chat Completions on bedrock-runtime and bedrock-mantle;
• Converse support on bedrock-runtime;
• implicit and explicit prompt caching;
• model-specific IDs under the openai. prefix;
• in-Region access on bedrock-mantle plus documented Geo and Global inference profile IDs on bedrock-runtime.

Because the model family is shared, a tier router can reuse much of the workload evaluation. It still needs explicit endpoint and API metadata: bedrock-runtime Responses is synchronous and does not provide server-side tools, while bedrock-mantle Responses supports server-side tools and asynchronous inference.

The endpoint and Region difference

Tier	bedrock-mantle in-Region	bedrock-runtime Geo	bedrock-runtime Global
Sol	us-east-1, us-east-2	US	Supported
Terra	us-east-1, us-east-2, us-west-2	US and India	Supported
Luna	us-east-1, us-east-2, us-west-2	US and India	Supported

If a workload must be processed only in Oregon, the current in-Region table rules out Sol before quality testing. A US Geo profile callable from Oregon is not equivalent: it can process the request in another US destination Region. Global routing expands the processing geography further and currently carries lower Standard token rates, but it is unsuitable when policy requires a bounded geography.

For bedrock-runtime, copy the exact model-specific inference profile ID and allowed source Region from the current model card. A Global profile’s destination set can expand over time; a geography-scoped profile keeps its existing destination set, but new profile IDs can be introduced. Check IAM and SCP access to all destinations before rollout.

See the full 1M context and Regions guide.

Choose Sol when failure is expensive

Start with Sol for tasks where stronger reasoning can materially improve the final outcome: complex code changes, ambiguous root-cause analysis, multi-step tool planning, or decisions whose review cost dominates inference cost.

Sol is not automatically the safest choice. Tool permissions, source trust, output validation, and human approval remain application responsibilities. It also has the narrowest documented Region coverage of the three.

Choose Terra for the production baseline

Terra is the practical first candidate when the workload mixes ordinary reasoning, tool use, text transformation, and structured output. It is the middle tier, so it can be a useful baseline for comparing both directions:

• test Sol only on cases where Terra misses the quality contract;
• test Luna on high-volume cases where Terra may provide more capability than needed.

Do not call Terra “best” without a workload-specific evaluation. The valuable property is that it creates a neutral baseline for a tiered routing decision.

Choose Luna for bounded, repeatable work

Luna is positioned for high-volume classification, summarization, routing, and real-time applications. It is a strong candidate when the input and output contract is narrow, errors are detectable, and latency or unit cost is important.

Use a strict schema, confidence or abstention path, and a fallback for hard cases. A cheaper first pass only saves money if bad classifications do not create expensive downstream work.

A routing policy that stays auditable

Avoid asking a model to choose its own tier without an external policy. Start with deterministic rules based on the product contract:

Luna: bounded schema + high volume + low ambiguity
Terra: general production default + mixed reasoning/tools
Sol: approved hard-case queue + quality-first contract

Then record the selected tier, route reason, prompt version, evaluation cohort, token use, latency, and outcome. Review false escalations and false downgrades.

Do not put sensitive content into a stronger tier or different Region unless the original authorization and data policy allow it.

Compare cost without copying stale prices

AWS currently lists the same Standard token rates for in-Region and Geo inference, with lower Standard rates for Global inference. Rates and service-tier support can change, so read the live model card and Bedrock pricing page at decision time.

Compare:

• uncached and cached input;
• cache-write premiums;
• output and reasoning tokens;
• retries and fallback calls;
• latency-related infrastructure cost;
• human review and correction time;
• cost per successful task.

The prompt caching guide provides a normalized worksheet for repeat prefixes.

Evaluation matrix

Use the same representative cases for each eligible tier. Score:

1. task completion and factual correctness;
2. required citation or evidence use;
3. JSON schema and parser success;
4. tool selection, arguments, retries, and duplicate actions;
5. latency distribution and timeout rate;
6. input, cache, reasoning, and output tokens;
7. cost per successful task;
8. long-context retrieval and conflict handling;
9. refusal, escalation, and human-review behavior.

Promote a cheaper tier only when it meets the contract. Escalate to Sol only for a defined failure class, not whenever a prompt contains the word “complex.”

Frequently asked questions

What is the difference between GPT-5.6 Sol, Terra, and Luna on Bedrock?

AWS positions Sol for the most complex reasoning and agentic coding, Terra for balanced everyday production work, and Luna for fast, high-volume work such as classification and summarization.

Do all three GPT-5.6 models have 1 million-token context on Bedrock?

Yes. AWS announced 1 million-token context support for Sol, Terra, and Luna on Amazon Bedrock on August 3, 2026.

Should I use GPT-5.6 Sol for every difficult workload?

No. Begin with a role-based candidate, then compare tiers on representative cases. Use the least costly and fastest configuration that reliably meets the task, tool, schema, latency, and safety contract.

Official sources

• AWS: Get started with GPT-5.6 Sol, Terra, and Luna
• AWS: GPT-5.6 1 million-token context announcement
• Amazon Bedrock model cards for OpenAI models
• Amazon Bedrock pricing
• AWS: APIs supported by Amazon Bedrock
• AWS: Route requests with cross-Region inference

Source check: August 19, 2026. Recheck model cards, endpoints, APIs, profile IDs, source and destination Regions, quotas, pricing, and feature support before selecting a tier.