I tried out OpenAI's gpt-oss-120b and gpt-oss-20b which became available on Amazon Bedrock

On August 5, 2025, OpenAI released open-weight language models "gpt-oss-120b" and "gpt-oss-20b". On the same day, it was announced that these models became available on AWS's Amazon Bedrock and Amazon SageMaker.

https://www.aboutamazon.jp/news/aws/openai-open-weight-models-now-available-on-aws

https://openai.com/ja-JP/index/introducing-gpt-oss/

In this article, I'll introduce the results of enabling access to these models on Amazon Bedrock in the North America Oregon region (us-west-2) and verifying their operation in the chat playground.

Model Access Settings

First, I selected "OpenAI" as the provider in the Bedrock dashboard for the Oregon region (us-west-2) and requested access to use "gpt-oss-120b" and "gpt-oss-20b".

I requested these models on the edit model access screen.

Using AWS CLI to search for models containing gpt-oss in the us-west-2 region, I obtained the following information:

$ aws bedrock list-foundation-models --region us-west-2 --query "modelSummaries[?contains(modelName, 'gpt-oss')]"  
[
    {
        "modelArn": "arn:aws:bedrock:us-west-2::foundation-model/openai.gpt-oss-120b-1:0",
        "modelId": "openai.gpt-oss-120b-1:0",
        "modelName": "gpt-oss-120b",
        "providerName": "OpenAI",
        "inputModalities": [
            "TEXT"
        ],
        "outputModalities": [
            "TEXT"
        ],
        "responseStreamingSupported": false,
        "customizationsSupported": [],
        "inferenceTypesSupported": [
            "ON_DEMAND"
        ],
        "modelLifecycle": {
            "status": "ACTIVE"
        }
    },
    {
        "modelArn": "arn:aws:bedrock:us-west-2::foundation-model/openai.gpt-oss-20b-1:0",
        "modelId": "openai.gpt-oss-20b-1:0",
        "modelName": "gpt-oss-20b",
        "providerName": "OpenAI",
        "inputModalities": [
            "TEXT"
        ],
        "outputModalities": [
            "TEXT"
        ],
        "responseStreamingSupported": false,
        "customizationsSupported": [],
        "inferenceTypesSupported": [
            "ON_DEMAND"
        ],
        "modelLifecycle": {
            "status": "ACTIVE"
        }
    }
]

From the CLI execution results, I found that at this point, the input and output are TEXT only, and "responseStreamingSupported": false indicates that streaming responses are not supported.

Operation Verification

I verified the operation of both models in the Bedrock chat playground.
I set the maximum output tokens to 8192 and ran in comparison mode with other settings at their default values.

I used the following prompt and compared the response content and processing time:

Create a simple explanation of the AWS shared responsibility model in Japanese that executives can intuitively understand.

As a result, although there was some distortion in the first heading (H2 element), I was able to get an appropriate response in Japanese Markdown format.

Price Comparison

I compared the billing unit price per output token between the main models available on Bedrock and the gpt-oss series.

• Unit price when using Bedrock in Oregon region on-demand

From this comparison, I found that gpt-oss-20b is priced almost the same as "Amazon Nova Lite," and the higher-end model gpt-oss-120b can be used at about half the price of "Claude 3 Haiku." These appear to be very strong options for cost-conscious use cases.

Throughput Performance Comparison

I also executed the AWS responsibility model prompt on other major models to measure the number of tokens and processing time. I compared the number of tokens that can be output per second (throughput).

I confirmed that gpt-oss-20b has throughput performance comparable to the fastest "Amazon Nova Micro," and gpt-oss-120b exceeds "Amazon Nova Pro." High performance can be expected not only in terms of price but also in response speed.

Comparison with Local LLM

For reference, I compared the throughput when running "gpt-oss-20b" on Amazon Bedrock versus a local environment (LM Studio on an M4 Pro Mac).

While there may be insufficient optimization settings in the local environment, I found that Amazon Bedrock provides excellent performance as an execution environment for gpt-oss-20b.

https://dev.classmethod.jp/articles/openai-gpt-oss-20b-mac-mini-m4-pro/

Summary

The newly released open-weight models "gpt-oss-20b" and "gpt-oss-120b" have been confirmed to be excellent choices in terms of both cost and performance.

These models could be effective candidates, particularly for tasks requiring low-cost and high-speed processing.

Currently, OpenAI models provided on Bedrock have limitations in regions and available functions, but I look forward to future service expansions.