I tried out the newly supported OpenAI GPT-5.6 Sol / Terra / Luna on Amazon Bedrock

Introduction

The GPT-5.6 family announced by OpenAI on July 9, 2026 became GA on Amazon Bedrock on July 13. It became available on Bedrock within just a few days of release.

https://aws.amazon.com/blogs/machine-learning/openai-gpt-5-6-sol-terra-and-luna-are-now-generally-available-on-amazon-bedrock/

https://openai.com/ja-JP/index/gpt-5-6/

With GPT-5.6, the naming convention has changed. The number 5.6 represents the generation, while Sol / Terra / Luna represent independently evolving tiers. The flagship Sol handles full-power reasoning, Terra is a balanced general-purpose tier, and Luna is the lightweight, low-cost option.

This is a comparison of the GPT-5.6 family and recent models as confirmed on July 14, 2026. Pricing is per 1M tokens (USD).

※ Context length is the value listed on the Amazon Bedrock model card. The official OpenAI documentation states a 1,050K context window, but it is 272K on Bedrock.

In Bedrock pricing, Sol is cheaper than GPT-5.5 for both Input and Output, and Terra is cheaper than GPT-5.4. Luna was newly added as the most affordable tier in the family.

Since the authentication and setup procedures are the same as in previous articles, please refer to those.

https://dev.classmethod.jp/articles/bedrock-openai-gpt55-gpt54-ga-responses-api/

Model IDs, pricing, and region support have been confirmed in the following official documentation.

• Model cards
• Pricing
• Region compatibility

Verification Details

Verification Environment

The GPT-5.6 family is exclusive to the bedrock-mantle endpoint, not the bedrock-runtime endpoint. It does not support the Invoke API / Converse API and is called via the OpenAI-compatible Responses API.

Here is the client initialization.

from openai import OpenAI
from aws_bedrock_token_generator import BedrockTokenGenerator

REGION = "us-east-1"
BASE_URL = f"https://bedrock-mantle.{REGION}.api.aws/openai/v1"

generator = BedrockTokenGenerator()
token = generator.get_token(credentials, REGION)

client = OpenAI(base_url=BASE_URL, api_key=token)

Non-Streaming Call

We sent the same prompt to all 3 models without streaming and confirmed that calls completed successfully. reasoning.effort is set to low for all.

response = client.responses.create(
    model="openai.gpt-5.6-sol",
    input="Amazon Bedrockとは何ですか？3文で簡潔に説明してください。",
    reasoning={"effort": "low"},
)

Here is a sample response from Sol.

Amazon Bedrock is a fully managed service provided by AWS for developing generative AI applications. It allows you to use foundation models from multiple leading companies via API, with the ability to select, customize, and evaluate models. You can build chatbots, text generation, search assistance, and more securely while minimizing infrastructure management.

All 3 models responded successfully. These are all single measurement values and should be treated as reference figures.

The initial latency for Terra and Luna was large at approximately 75 seconds, but since it dropped significantly on the second call, this is likely due to cold start effects. On the second call, Terra dropped to 1.12s and Luna to 0.57s. Sol was 2.22s from the first call, suggesting it may have already been in a warm state.

For a prompt simply asking for an explanation, all 3 models returned reasoning_tokens of 0.

Sol:
ResponseUsage(input_tokens=23, input_tokens_details=InputTokensDetails(cached_tokens=0, cache_write_tokens=0), output_tokens=102, output_tokens_details=OutputTokensDetails(reasoning_tokens=0), total_tokens=125)

Terra:
ResponseUsage(input_tokens=23, input_tokens_details=InputTokensDetails(cached_tokens=0, cache_write_tokens=21), output_tokens=107, output_tokens_details=OutputTokensDetails(reasoning_tokens=0), total_tokens=130)

Luna:
ResponseUsage(input_tokens=23, input_tokens_details=InputTokensDetails(cached_tokens=0, cache_write_tokens=21), output_tokens=115, output_tokens_details=OutputTokensDetails(reasoning_tokens=0), total_tokens=138)

Streaming Call

We called the same prompt with stream=True and confirmed the streaming response.

stream = client.responses.create(
    model="openai.gpt-5.6-sol",
    input="Amazon Bedrockとは何ですか？3文で簡潔に説明してください。",
    reasoning={"effort": "low"},
    stream=True,
)

Results measured in a warm state. TTFT is the time until the first response.output_text.delta event is received.

The event structure was common across all 3 models. It starts with response.created, followed by repeated response.output_text.delta events, and ends with response.completed.

Verifying reasoning.effort

We called Sol while varying reasoning.effort and checked the increase or decrease in reasoning_tokens. The prompt was "Please prove that there are infinitely many prime numbers." We only checked whether it works and the token count, without evaluating accuracy.

response = client.responses.create(
    model="openai.gpt-5.6-sol",
    input="素数が無限に存在することを証明してください。",
    reasoning={"effort": "max"},
)

While the simple explanation prompt in the non-streaming case returned reasoning_tokens=0, the proof prompt returned reasoning tokens. From low to high, reasoning_tokens remained roughly flat at 33–37, with no clear increase by effort level. Latency also reversed from low 3.34s → high 2.50s, and since these are all single measurement values, they fall within the margin of measurement error. Only with max did reasoning_tokens clearly increase to 104.

Specifying reasoning.effort="ultra" returned an error.

Invalid value: 'ultra'. Supported values are: 'none', 'minimal', 'low', 'medium', 'high', 'xhigh', and 'max'.

The valid values for reasoning.effort on Bedrock Mantle are 7 levels: none / minimal / low / medium / high / xhigh / max. ultra was not accepted at the time of this verification.

Confirming Chat Completions API Is Not Supported

At least at the time of this verification, Bedrock Mantle does not support the Chat Completions API, and calling it returned the following error.

client.chat.completions.create(
    model="openai.gpt-5.6-sol",
    messages=[{"role": "user", "content": "Hello"}],
)

400 - {'error': {'code': 'validation_error', 'message': "The model 'openai.gpt-5.6-sol' does not support the '/v1/chat/completions' API", 'param': None, 'type': 'invalid_request_error'}}

Summary

We called GPT-5.6 Sol / Terra / Luna, which became GA on Amazon Bedrock, via the Responses API at the Bedrock Mantle endpoint. All 3 models responded successfully in both non-streaming and streaming modes, confirming that the GPT-5.6 family can be selected according to use case on Bedrock.

In Bedrock pricing, Sol is cheaper than GPT-5.5 and Terra is cheaper than GPT-5.4. Luna has been added as a low-cost tier at $1/$6, making it a prime candidate to try first for lightweight tasks such as summarization and classification.

At the time of this verification, only the Standard Service Tier was available, with Priority / Flex / Reserved and cross-region inference not yet supported (In-Region only). Image input is supported but was not verified this time.

As a note of caution, the model card states that traffic data flagged by the classifier for abuse detection may be retained for up to 30 days. Since this is an exception to the standard Bedrock zero data retention policy, it is advisable to review the model card before putting it into production use.