AWS Bedrock GPT-5.6: 1M Context and Regions

Quick answer

On Amazon Bedrock, GPT-5.6 Sol, Terra, and Luna each support a 1 million-token context window. Region behavior now depends on the endpoint and inference option:

Bedrock model	bedrock-mantle in-Region ID	bedrock-runtime Geo IDs	bedrock-runtime Global ID
GPT-5.6 Sol	openai.gpt-5.6-sol	us.openai.gpt-5.6-sol	global.openai.gpt-5.6-sol
GPT-5.6 Terra	openai.gpt-5.6-terra	us.openai.gpt-5.6-terra, in.openai.gpt-5.6-terra	global.openai.gpt-5.6-terra
GPT-5.6 Luna	openai.gpt-5.6-luna	us.openai.gpt-5.6-luna, in.openai.gpt-5.6-luna	global.openai.gpt-5.6-luna

All three support Responses, Chat Completions, and Converse on bedrock-runtime; Responses and Chat Completions are OpenAI-compatible, while Converse uses the AWS SDK request shape. The existing bedrock-mantle path remains available for in-Region Responses and server-side tools. The endpoint families are not feature-equivalent.

For setup code and authentication, use the Bedrock GPT-5.6 Responses API guide.

What the 1M limit changes

The expanded window makes it technically possible to send larger repositories, document sets, or agent histories in one request. It does not prove that sending everything is the best design.

Before using the full window, test:

1. the rendered token count, including system instructions, tools, images, files, and prior outputs;
2. whether the required evidence is still retrieved and cited when surrounded by irrelevant material;
3. first-token and end-to-end latency at typical and worst-case sizes;
4. input, cache-write, cache-read, reasoning, and output token cost;
5. failure and retry behavior near the limit;
6. whether retention, logging, and data-location controls permit the full source set.

A 1M capacity can reduce application-side chunking, but retrieval is still useful when it narrows evidence, enforces permissions, or keeps changing sources current. Summarization and compaction are also still useful when an agent history contains stale decisions or verbose tool results.

Choose in-Region, Geo, or Global deliberately

Use bedrock-mantle when the request must use a documented single-Region endpoint or needs its server-side tool surface. Its current in-Region model-card coverage is us-east-1 and us-east-2 for Sol, with us-west-2 also available for Terra and Luna.

Use bedrock-runtime when you need Responses, Chat Completions, or Converse with a Geo or Global inference profile. AWS currently documents no in-Region GPT-5.6 option on this endpoint: the request must name a model-specific Geo or Global inference profile ID. Geo keeps processing within the named geography, but prompts and outputs can move outside the source Region to another destination Region in that geography. Global can route to supported commercial Regions worldwide, and its destination set can expand over time.

This makes the source Region, processing geography, and storage behavior separate controls. A Geo label is not a promise that processing stays in the source Region, and Global is not appropriate when workload policy requires a bounded geography.

Choose a Region only after checking:

• model availability for the selected endpoint, API, source Region, and profile;
• the location of source data, vector stores, tools, logs, and dependent services;
• residency, abuse-detection storage, and contractual requirements;
• IAM and SCP access to every profile destination Region;
• in-Region versus cross-Region quotas and live pricing;
• failover behavior if the chosen Region or model is unavailable.

The two endpoint patterns are:

in-Region:    https://bedrock-mantle.us-east-1.api.aws/openai/v1
cross-Region: https://bedrock-runtime.us-east-1.amazonaws.com/openai/v1

On bedrock-runtime, use the Geo or Global inference profile ID as the request model. Changing endpoint, source Region, or profile is an application and operations change, not just a console selection. Revalidate credentials, network policy, SCPs, quotas, logs, pricing, and downstream data flows.

Keep cross-Region operations observable

CloudWatch and CloudTrail stay centered in the source Region. AWS documents additionalEventData.inferenceRegion in CloudTrail for identifying the Region that processed a cross-Region request. Record the endpoint, API, source Region, inference profile ID, request ID, processing Region when available, tokens, latency, and errors without logging secrets or sensitive prompt bodies.

Use the AWS/Bedrock metrics for bedrock-runtime traffic and the separate mantle monitoring surface for bedrock-mantle. Do not merge both into one dashboard without an endpoint dimension.

Avoid three context mistakes

Do not mix provider limits

The Bedrock announcement and model cards are the authority for Bedrock access. Direct OpenAI model documentation can show different context limits, model aliases, features, or availability. Keep provider metadata separate in model registries and user interfaces.

Do not treat capacity as guaranteed answer quality

Long input can contain conflicts, duplicates, malicious instructions, or stale files. Preserve source boundaries, prioritize trusted evidence, and evaluate whether the answer cites the correct portion of the input.

Do not resend stable context blindly

If a large prefix repeats, measure prompt caching. Bedrock GPT-5.6 supports implicit caching by default and explicit breakpoints for controlled stable prefixes. See the prompt caching cost guide.

Deployment checklist

• Select Sol, Terra, or Luna from the workload, not only the window size.
• Confirm the exact endpoint, API, source Region, and model or inference profile ID in configuration.
• Use bedrock-mantle for documented in-Region access or its server-side tools; use bedrock-runtime for Geo or Global routing.
• Allow every profile destination Region in IAM and SCP policy, then test a denied destination.
• Measure typical, p95, and worst-case token counts.
• Run evidence-retrieval and conflicting-source evaluations at long context.
• Log endpoint, API, source Region, profile, processing Region when available, token categories, latency, and errors.
• Test quota exhaustion, timeout, retry, and Region recovery.
• Recheck the live model card before production rollout.

Frequently asked questions

Which GPT-5.6 models have a 1 million-token context window on Amazon Bedrock?

AWS states that GPT-5.6 Sol, Terra, and Luna each support a 1 million-token context window on Amazon Bedrock.

Which AWS Regions support GPT-5.6 on Bedrock?

It depends on the endpoint and inference option. bedrock-mantle provides documented in-Region access, while bedrock-runtime uses model-specific Geo or Global inference profile IDs. Check the current model card for allowed source and destination Regions before deployment.

Does a 1 million-token context window remove the need for retrieval or caching?

No. A large window is a capacity limit, not a relevance, latency, or cost guarantee. Measure the real token count, keep evidence scoped, test answer quality, and cache a stable repeated prefix when the workload supports it.

Official sources

• AWS: GPT-5.6 Sol, Terra, and Luna now support 1 million-token context windows
• AWS model card: GPT-5.6 Sol
• AWS model card: GPT-5.6 Terra
• AWS model card: GPT-5.6 Luna
• AWS: Route requests with cross-Region inference
• AWS: APIs supported by Amazon Bedrock

Source check: August 19, 2026. Recheck context, endpoint features, profile IDs, source and destination Regions, quotas, pricing, and logging before deployment.