Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- AWS scopes support to OpenAI models on Bedrock in Regions where Bedrock offers OpenAI GPT models; this does not mean every model, Region, or account is enabled.
- Explicitly set reasoning.summary to auto in a Responses request; summary detail can vary by model.
- Read the summary array from the output item whose type is reasoning; the final answer is a separate output.
- The announcement does not list a uniform summary surcharge or account quota. Check the target model, Region, permissions, and actual billing before rollout.
The short answer: request a summary, then read it from Responses output
Amazon Bedrock supports reasoning.summary for OpenAI models through the Responses API. Explicitly request a summary, then find the output item whose type is reasoning and read its summary array. Handle that data separately from the final answer.
AWS announced the capability on October 9, 2026. It applies to OpenAI models on Bedrock in AWS Regions where Bedrock offers OpenAI GPT models. The notice does not extend the feature to every Bedrock model or confirm access for a particular account. For separate model, endpoint and regional coverage, see the English Bedrock OpenAI Runtime guide under Further reading.
Use the Bedrock Runtime endpoint and credentials
The Amazon Bedrock Quickstart recommends bedrock-runtime for new applications. Its OpenAI-compatible Responses example uses https://bedrock-runtime.<region>.amazonaws.com/openai/v1; replace region with the AWS Region you selected. This is not the OpenAI platform api.openai.com endpoint.
The Quickstart example puts a Bedrock-generated short-term API key in a variable named OPENAI_API_KEY. The variable name does not mean the key came from OpenAI. For production, follow AWS guidance for IAM roles or temporary credentials, and keep credentials on the server rather than in browser code, public logs, or source control.
Explicitly request a summary in the Responses call
Use the Responses API and set reasoning.summary to auto, for example {"reasoning":{"summary":"auto"}}. This explicitly asks the model to return a readable summary. If the request omits the field, the application should not expect one.
auto selects the most detailed summary mode supported by the model in use, but supported options and returned content can differ. OpenAI's reasoning guide also says summaries must be explicitly requested; do not assume one model or interface's behavior applies to another.
Read the summary separately from the final answer
Inspect the response's output items, locate the item with type reasoning, and read its summary array. Summary entries commonly contain a type and text; parse the returned structure according to the current model documentation. Reading only the convenience field output_text can miss the summary.
A reasoning summary describes the model's approach in readable form; it is not a verbatim internal reasoning trace. OpenAI says raw reasoning tokens are not exposed. A summary can help with explanation or debugging, but it does not prove an answer is correct; verify facts and results independently.
Check model, Region, and data-residency eligibility
AWS says support covers OpenAI models on Bedrock in the AWS Regions where OpenAI GPT models are available, including in-Region, geographic (GEO) cross-Region, and global cross-Region inference. Availability is conditional on a model being offered in that Region; it does not establish enablement for every Region, model, or account.
The OpenAI-compatible example in the Bedrock Quickstart uses a global system inference profile. If your workload has residency requirements, select a supported geographic inference profile from the model card and verify it in the target account and Region. The announcement gives no uniform summary surcharge, exact quota, or account-activation status.
Troubleshoot missing summaries from the request and response
Confirm that the call uses the Responses API and actually includes reasoning.summary=auto. Then check that your code scans the full output array instead of reading only output_text. After that, verify the Bedrock model identifier, Region, inference profile, and account access; an empty summary alone does not prove a service fault.
Before integration, use a non-sensitive request in the target account to check response structure, access, cost and routing, then decide how to display the summary. A documented parameter alone does not establish that a particular model has returned one.
Sources
Frequently Asked Questions
Is a reasoning summary the complete chain of thought?
No. It is a readable summary returned alongside the answer. OpenAI says raw reasoning tokens are not exposed, so the summary is neither a verbatim internal trace nor proof that the answer is correct.
Can I use reasoning.summary with Chat Completions or Converse too?
Do not infer that from this announcement. AWS specifies reasoning.summary through the Responses API. Check Bedrock documentation for each other interface rather than copying the same field across APIs.
Does this work in every AWS Region and account without extra cost?
The announcement limits regional availability to Regions where Bedrock offers OpenAI GPT models. It does not confirm access for an individual account or publish a uniform summary surcharge or quota. Check the target model, Region, account, and current billing.