OpenAI introduces gpt-oss, its open AI models

OpenAI has introduced gpt-oss-120b and gpt-oss-20b, two reasoning models with open weights distributed under the Apache 2.0 license. The company is targeting developers who want to adapt AI and use it in their own systems, including agents that can browse the web or run Python code.

Both models are text-only and follow instructions. They are also compatible with OpenAI's Responses API, the interface for building applications that combine a model with external tools and multistep workflows.

The main difference is their size: gpt-oss-120b has 120 billion parameters, while gpt-oss-20b has 20 billion. Parameters are the internal values a model uses to learn patterns and generate responses. In general, a larger model needs more resources, while a smaller one may be more practical for certain systems and deployments.

What they can do

OpenAI designed these models for agent workflows. That means they do more than answer a question: they can follow instructions, decide when to use a tool and complete a task in multiple steps.

Their features include:

• Use of tools such as web searches and Python code execution.
• Adjustable reasoning effort based on the complexity of the task.
• Customization by developers.
• Generation of structured outputs, which is useful for integrating them into applications.
• Compatibility with the Responses API.

Controlling reasoning effort can be useful in practice. Classifying hundreds of messages, for example, does not always require the same amount of time as analyzing a contract or writing code. Developers can choose a balance between quality, speed and resource use.

Because they have open weights, the models offer more room for adaptation than a closed service. A company can adjust their behavior for a specific use case or integrate them into a system under its own conditions. This also means OpenAI will not be able to centrally apply new safety measures after distribution.

The safety warning

OpenAI acknowledges that open models have a different risk profile from models used only through an API. A determined attacker could modify them to bypass their safety refusals or optimize them for harmful purposes, without OpenAI being able to revoke access or impose a later mitigation.

That is why the company warns that some developers and businesses will need to add their own safeguards. A model may follow OpenAI's safety policies by default, but the overall system will also depend on who integrates it, which tools they connect to it and what controls they apply.

What the tests show

OpenAI evaluated gpt-oss-120b in three areas from its preparedness framework: biological and chemical risks, cybersecurity and the autonomous improvement of AI systems. According to the company, the model in its default configuration did not reach its indicative high-capability thresholds in any of them.

The company also tried to increase its capabilities through adversarial training, a form of training designed to simulate what an attacker might do. Even after that process, OpenAI says gpt-oss-120b did not reach the high level defined for biological and chemical risks or cyber risks.

It also compared the model with other open models and concluded that, in most tests, at least one model already available came close to the performance of adversarially tuned gpt-oss-120b. OpenAI's interpretation is that the release does not represent a significant advance in the frontier of biological capabilities among open models.

For you, the impact will depend on whether you develop or use AI software: there will be two OpenAI models that can be adapted more freely, but the responsibility for keeping them safe will be much greater outside the company's managed products and APIs. What matters now is seeing what applications the community builds and what safeguards accompany each one.