GPT-5.6 Inference Spans 25 AWS Regions on Bedrock

AWS and OpenAI added cross-Region inference for the GPT-5.6 model family to Amazon Bedrock on August 20, 2026. The deployment covers more than 25 AWS Regions and dynamically routes requests toward available capacity, reducing localized throttling during traffic bursts.

For teams operating production agents, the release moves capacity management into the inference layer. Applications can target a logical profile instead of binding every request to one AWS Region, while geographic profiles retain residency boundaries.

GPT-5.6 models and pricing

The launch includes three general-purpose variants. GPT-5.6 Sol targets long-horizon agentic coding, cyber defense, and complex scientific research. GPT-5.6 Terra handles production workloads such as code generation, structured extraction, and content processing at approximately half Sol’s operational cost. GPT-5.6 Luna prioritizes throughput and low latency.

All three models provide a 1,000,000-token context window, accept text and image inputs, and return text. Native reasoning, server-side tool calling, and prompt caching are available across the family. Cached input tokens receive discounts of up to 90%, which makes repeated long-context agent workflows materially cheaper. Teams evaluating context windows should account for both the maximum window and the cost of repeatedly sending large prefixes.

Model	Standard or geographic input	Standard or geographic output	Global input	Global output
GPT-5.6 Sol	$5.50 / 1M tokens	$33.00 / 1M tokens	$5.00 / 1M tokens	$30.00 / 1M tokens
GPT-5.6 Terra	$2.20 / 1M tokens	$13.20 / 1M tokens	$2.00 / 1M tokens	$12.00 / 1M tokens
GPT-5.6 Luna	$0.22 / 1M tokens	$1.32 / 1M tokens	$0.20 / 1M tokens	$1.20 / 1M tokens

Global profiles receive an approximately 10% token discount through broader resource optimization. Usage also counts toward existing AWS customer commitments, including EDP agreements. The pricing difference makes global routing attractive for workloads without residency constraints, while geographic routing is the safer default for regulated data.

Geographic and global profiles

Geographic inference profiles route requests among defined regional boundaries. A profile such as us.openai.gpt-5.6-terra can distribute execution across US East N. Virginia, US East Ohio, and US West Oregon while keeping processing inside the specified geography.

Global inference profiles use identifiers such as global.openai.gpt-5.6-sol and can route across supported commercial AWS Regions worldwide according to real-time hardware capacity. This architecture helps absorb sudden demand without requiring application-level failover logic, but global routing should be selected only after confirming that data handling policies permit it.

Requests run through the Amazon Bedrock bedrock-runtime endpoint. Developers can use the OpenAI Responses API at https://bedrock-runtime.{region}.amazonaws.com/openai/v1, OpenAI Chat Completions API compatibility, or Bedrock’s Converse and ConverseStream APIs. Existing teams building streaming LLM applications can use ConverseStream without adopting a separate provider endpoint.

Governance and production operations

IAM credentials provide the standard authentication path. Temporary bearer tokens generated with the aws-bedrock-token-generator Python package can also produce API keys valid for up to 12 hours.

Prompts and completions are not shared with OpenAI, are not retained for model training, and have zero operator access at the hardware chip level. Bedrock can log invocation payloads to Amazon S3 or CloudWatch Logs, while CloudWatch exposes token counts, latency, throttles, and error rates. Cost Explorer and the AWS Cost and Usage Report itemize spend by model ID and profile ID, giving AI observability systems useful routing and cost dimensions.

If you deploy GPT-5.6 on Bedrock, choose a geographic profile when residency is a hard requirement and a global profile when capacity and the 10% discount matter more. Test throttle rates, tail latency, cache behavior, and regional policy constraints before switching production traffic.