I’ve spent years tinkering with AI, and one of my biggest frustrations has always been the “black box” problem. You feed a prompt into a powerful model, get a great answer, but you have very little say in how it works under the hood. You can’t really customize it for your specific needs, and you’re locked into whatever the provider gives you. It’s like having a supercar but not being allowed to pop the hood.
Well, that all just changed. In a move that’s honestly a massive deal for developers and businesses, AWS has brought two of OpenAI’s new open-weight models to the cloud. This is the moment a lot of us have been waiting for.
Say hello to gpt-oss-120b and gpt-oss-20b, now available on Amazon Bedrock and Amazon SageMaker JumpStart. The “oss” stands for open-source software, and the “open weight” part is the magic here. It means you get unprecedented access to the model’s inner workings. You can modify them, fine-tune them, and build highly specialized AI that’s perfectly tailored to your world.
This is not just another model release; it’s a fundamental shift in how we can build with AI.
These models are absolute powerhouses, especially for complex tasks like coding, deep scientific analysis, and mathematical reasoning. They’re right up there with the best of them in terms of performance. And the specs? They are wild.
Both models come with a jaw-dropping 128K context window. To put that in perspective, that’s enough to feed the model an entire technical manual, a massive codebase, or a very, very long legal document and have it reason over the entire thing in one go. The days of awkwardly chunking up your documents are numbered!
Plus, they come with adjustable reasoning levels. You can literally tell the model to use “low,” “medium,” or “high” reasoning effort depending on your task. Need a quick answer? Go low. Need it to solve a complex multi-step problem? Crank it to high. It’s like having a throttle for the model’s brain.
So, how do you get your hands on this awesome new tech? AWS gives you two clear paths, depending on your style.
⚙️ Route 1: The Amazon Bedrock Way (The Easy Button)
If you want power without the headache of managing infrastructure, Amazon Bedrock is your best friend. It’s a serverless platform, which means you can just call the model via an API and AWS handles all the scaling and management behind the scenes. It’s the fastest way to get up and running.
Here’s how simple it is:
- Request Access: Log into your AWS account, head to the Amazon Bedrock console, and find the “Model access” section. Find the new OpenAI models and click to request access.
- Test in the Playground: Once you’re approved (it’s usually fast), you can immediately start experimenting in the Chat playground. This is perfect for getting a feel for the model’s capabilities and testing prompts, like the financial planning example from the announcement.
- Integrate with Your Code: This is the super cool part. Bedrock provides an OpenAI-compatible endpoint. This means you can use the official OpenAI Python SDK, the one you probably already know and love, and just point it to Bedrock. You don’t have to rewrite a bunch of code. You just set a couple of environment variables, and you’re golden.
Here’s a quick code snippet to show you what I mean:
# First, set these in your terminal
export OPENAI_API_KEY="<your-bedrock-api-key>"
export OPENAI_BASE_URL="https://bedrock-runtime.us-west-2.amazonaws.com/openai/v1"
# Then, your Python code is standard OpenAI!
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
messages=[{ "role": "user", "content": "Tell me the square root of 42 ^ 3" }],
model="openai.gpt-oss-120b-1:0",
)
print(response.choices[0].message.content)It’s that easy. You get the power of OpenAI’s models with the security, scalability, and integration of the AWS ecosystem.
🚀 Route 2: The SageMaker JumpStart Path (For Power Users)
What if you want to go deeper? What if you want to fine-tune the model on your company’s private data to create a specialized expert? That’s where Amazon SageMaker JumpStart comes in.
SageMaker is AWS’s all-in-one machine learning platform. JumpStart is a feature that lets you, well, jumpstart your process with pre-trained models. It’s designed for ML practitioners who want maximum control.
Your workflow here looks a little different:
- Find the Model: Go to the SageMaker AI console and browse the JumpStart models. You’ll find
gpt-oss-120bandgpt-oss-20bwaiting for you. - Deploy with One Click: Select the model you want, choose an instance type (the virtual server it will run on), and click “Deploy.” SageMaker handles the complex setup and, after a few minutes, gives you a secure, private endpoint.
- Customize and Build: Now that you have a deployed model, you can use all the powerful tools in SageMaker to fine-tune it, evaluate its performance, and integrate it into your production ML pipelines. This is the path for building truly unique, proprietary AI capabilities.
✨ More Awesome Features to Know
Beyond just the access, there are a few other things that make this release so great.
- 💡 Full Transparency: Both models provide full chain-of-thought outputs. This means when you ask it a complex question, it doesn’t just give you the answer. It shows you its work: the step-by-step reasoning it used to get there. For any application where you need to trust and verify the output, this is an absolute game-changer.
- 🔧 Total Customization: I can’t stress this enough. “Open weights” means you have the freedom to adapt these models. You can build a specialized version for medical research, legal contract analysis, or even to mimic your company’s specific brand voice. The possibilities are endless.
- 🤝 Familiar Tools: The models use the standard GPT-4 tokenizer, so if you have tools or prompts built around that, they’ll work seamlessly. And as I mentioned, the Bedrock endpoint’s compatibility with the OpenAI SDK is a huge productivity win.
- ✍️ Ready for AI Agents: These models are designed to work with external tools, making them perfect for building AI agents. You can hook them up to a framework like Strands Agents, give them access to APIs or databases, and let them execute complex, multi-step tasks on their own. It’s the next frontier of automation.
Here’s a little taste of what that agentic code looks like with Strands:
from strands import Agent
from strands.models import BedrockModel
# Point Strands to the Bedrock model
bedrock_model = BedrockModel(
model_id="openai.gpt-oss-120b-1:0",
region_name="us-west-2"
)
agent = Agent(model=bedrock_model)
# Unleash the agent!
agent("Tell me the square root of 42 ^ 3")This is a massive leap forward. Giving developers this level of power and control, combined with the robustness of AWS, is going to unlock a whole new wave of AI innovation. Whether you’re a startup looking for an edge or a large enterprise wanting to build specialized AI, this is your moment.
Go check them out on the Amazon Bedrock or SageMaker AI consoles today. It’s time to pop the hood and start building.
This collaboration intensifies the “AI cloud wars” between AWS and Microsoft. While Microsoft’s Azure has long held exclusive rights to resell OpenAI’s flagship proprietary models like GPT-4, this deal sidesteps that exclusivity. The new gpt-oss-120b and gpt-oss-20b models are “open-weight” and released under the Apache 2.0 license, which allows for broad commercial use and redistribution, enabling AWS to offer them to its vast customer base.
For developers and enterprises, this move provides several key advantages:
- Greater Choice and Flexibility: Users are no longer limited to a single cloud provider for OpenAI technology and can now select from a wider array of models on the AWS platform, tailoring their choice to specific needs and budgets.
- Enhanced Control and Customization: The open-weight nature of the models allows organizations to fine-tune them on their own data and deploy them within their secure AWS environment, maintaining full control over their infrastructure.
- Cost-Performance: AWS highlights that the larger
gpt-oss-120bmodel is significantly more price-performant than comparable competing models, potentially lowering the barrier to entry for deploying large-scale AI.