I tried out reasoning summaries now available on OpenAI models in Bedrock
This page has been translated by machine translation. View original
Introduction
On October 9, 2026, OpenAI models on Amazon Bedrock added support for reasoning.summary in the Responses API, making it possible to receive reasoning summaries along with responses.
This support for reasoning summaries is an update that improves compatibility with the OpenAI API for the bedrock-runtime endpoint, which added support for OpenAI GPT models in August 2026.
This article introduces the results of calling GPT-6 Luna using the Responses API via curl (SigV4 signing) and the Converse API via AWS CLI, receiving reasoning summaries in both cases.
Receiving Summaries with the Responses API
The steps for calling OpenAI models on Bedrock, including inference profiles and authentication, are explained in the following article.
We call the model with model ID global.openai.gpt-6-luna in us-east-1. The request body is saved as req-luna-auto-hard.json.
{
"model": "global.openai.gpt-6-luna",
"input": "1から100までの素数の個数と、その総和を求めてください。",
"reasoning": {
"effort": "medium",
"summary": "auto"
},
"store": false
}
Since bedrock-runtime in aws-cli 2.37.12 does not have a subcommand for calling the Responses API, we sign and call it using curl's --aws-sigv4 option. Credentials and region are passed via the environment variables AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_SESSION_TOKEN, and AWS_REGION (us-east-1).
curl -sS --aws-sigv4 "aws:amz:${AWS_REGION}:bedrock" \
--user "${AWS_ACCESS_KEY_ID}:${AWS_SECRET_ACCESS_KEY}" \
-H "x-amz-security-token: ${AWS_SESSION_TOKEN}" \
-H 'Content-Type: application/json' \
-d @req-luna-auto-hard.json \
"https://bedrock-runtime.${AWS_REGION}.amazonaws.com/openai/v1/responses"
An excerpt of the output from the response is shown below. The encrypted_content is redacted, leaving only its length.
"output": [
{
"content": [],
"encrypted_content": "<redacted len=1228>",
"summary": [
{
"text": "**Calculating prime statistics**\n\nI'm preparing the response in Japanese and checking both requested values carefully. The prime numbers from 1 through 100 total 25, remembering that 1 is not prime. I'm also verifying their sum in groups to avoid an arithmetic slip: the total comes to 1,060. I'll provide the complete list so the result is transparent, followed by the count and sum. The answer should stay concise while still explaining the key point about 1.",
"type": "summary_text"
}
],
"type": "reasoning"
},
{
"content": [
{
"text": "1は素数ではありません。1から100までの素数は**25個**で、総和は**1060**です。",
"type": "output_text"
}
],
"role": "assistant",
"type": "message"
}
]
The values that can be specified for summary vary by model. Per-model support is documented in OpenAI's reasoning guide. For Luna, auto, concise, and detailed could be specified. auto is the value that selects the most detailed summary available for that model.
How summary Specification Affects the Returned Summary
When the prime number question from earlier was sent twice each with auto, concise, and detailed specified for summary, all six calls returned a summary in the summary array. Calls made without specifying a value returned an empty summary array.
On the other hand, when asking "What is 17 * 24? Please answer while explaining your thought process along the way.", 2 out of 8 calls with auto or concise specified returned an empty summary array.
For the prime number question asked in Japanese, the responses were in Japanese, but all six summaries—both headings and body text—were in English.
Streaming
By adding "stream": true to the request body and calling curl with the -N flag, the summary arrived before the response. The summary text streamed via response.reasoning_summary_text.delta events, and the response text via response.output_text.delta events.
Receiving Summaries with the Converse API
With the Converse API as well, passing reasoning that includes summary in additionalModelRequestFields allowed summaries to be retrieved.
The request is saved as req-converse-luna.json.
{
"modelId": "global.openai.gpt-6-luna",
"messages": [{"role": "user", "content": [{"text": "1から100までの素数の個数と、その総和を求めてください。"}]}],
"additionalModelRequestFields": {"reasoning": {"effort": "medium", "summary": "auto"}}
}
Call it with the following command.
aws bedrock-runtime converse --region us-east-1 --cli-input-json file://req-converse-luna.json
An excerpt of the output from the response is shown below.
"output": {
"message": {
"role": "assistant",
"content": [
{
"reasoningContent": {
"reasoningText": {
"text": "**Calculating prime totals**\n\nI'm answering in Japanese and providing the prime numbers below 100. I'm checking the count and sum carefully: there are 25 primes, and their total is 1,060. I'll include the complete list so the result is easy to verify, followed by the count and sum. Keeping the explanation concise should fit the likely goal while still showing how the totals were obtained."
}
}
},
{
"text": "1から100までの素数は **25個**、その総和は **1060** です。\n\n素数は \n2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53, 59, 61, 67, 71, 73, 79, 83, 89, 97 \nです。"
}
]
}
}
For calls made without specifying summary, reasoningContent contained redactedContent with unreadable content instead of a summary.
Summary
OpenAI models on Amazon Bedrock can now receive reasoning summaries.
This is expected to resolve errors that had been occurring because bedrock-runtime previously did not support reasoning summaries, and to eliminate the need for workaround patches.
Please make use of the reasoning summaries now available on Bedrock as one tool for adjusting the response quality of GPT models.
