Skip to content
gpt-oss-120b & gpt-oss-20b: Apache 2.0 Open-Weight Reasoning Models
Source: openai.com

gpt-oss-120b & gpt-oss-20b: Apache 2.0 Open-Weight Reasoning Models

Openai Policy Open Source

Sources: https://openai.com/index/gpt-oss-model-card, openai.com

Share on X Share on LinkedIn

TL;DR

  • gpt-oss-120b and gpt-oss-20b are open-weight reasoning models released under the Apache 2.0 license and the gpt-oss usage policy. OpenAI
  • They are text-only, compatible with the Responses API, and designed to be used within agentic workflows with strong instruction following, tool use (web search and Python code execution), and reasoning capabilities—including the ability to adjust the reasoning effort for tasks that don’t require complex reasoning. OpenAI
  • The models are customizable, provide full chain-of-thought (CoT), and support Structured Outputs. Safety is foundational to our approach to open models. OpenAI
  • They present a different risk profile than proprietary models: once released, determined attackers could fine-tune them to bypass safety refusals or directly optimize for harm without the possibility for OpenAI to implement additional mitigations or to revoke access. Extra safeguards may be needed by developers and enterprises to replicate system-level protections built into API-served models. OpenAI
  • Scalable capability evaluations on gpt-oss-120b showed the default model does not reach the indicative thresholds for High capability in any of the three Tracked Categories of the Preparedness Framework (Biological and Chemical capability, Cyber capability, and AI Self-Improvement). OpenAI
  • OpenAI reaffirms commitment to beneficial AI and raising safety standards across the ecosystem. Publication Aug 22, 2025; Safety Aug 7, 2025; Publication Aug 7, 2025. OpenAI

Context and background

OpenAI introduced two open-weight reasoning models, gpt-oss-120b and gpt-oss-20b, to offer open-weight options under the Apache 2.0 license and the gpt-oss usage policy. These text-only models were developed with feedback from the open-source community and are designed to be used within a broad range of systems created and maintained by diverse stakeholders. They are intended to be compatible with OpenAI’s Responses API and are built to operate within agentic workflows that demand reliable instruction following, tool use such as web search and Python code execution, and flexible reasoning capabilities. OpenAI The model card framing emphasizes safety as foundational to open models. It notes that open models present a different risk profile from proprietary models because released models could be fine-tuned to bypass safety refusals or optimize for harm. Therefore, OpenAI indicates that developers and enterprises may need to implement additional safeguards to replicate the system-level protections embedded in models served through its API and products. The document specifically distinguishes a model card from a system card, reflecting that gpt-oss models will be used as part of many different systems built by many stakeholders. OpenAI In evaluating safety and capabilities, OpenAI reports conducting scalable capability evaluations on gpt-oss-120b and confirming that the default model does not reach High capability thresholds in the three tracked categories of their Preparedness Framework: Biological and Chemical capability, Cyber capability, and AI Self-Improvement. The company frames this as part of its ongoing commitment to beneficial AI and to raising safety standards across the ecosystem. Publication and safety milestones associated with these findings are noted as Aug 22, 2025 and Aug 7, 2025. OpenAI

What’s new

The release covers two new open-weight models: gpt-oss-120b and gpt-oss-20b. Both models are text-only and offered under the Apache 2.0 license, alongside OpenAI’s gpt-oss usage policy. They are designed to be used with the Responses API and to function within agentic workflows that require robust instruction following, tool usage (including web search and Python code execution), and reasoning that can be tuned for lighter or heavier cognitive effort as task demands dictate. The models are customizable, provide full chain-of-thought (CoT), and support Structured Outputs. A key emphasis is safety, with explicit recognition of the different risk profile for open models and a caution that attackers could attempt to circumvent safeguards if the models are released without additional protections. OpenAI In addition, OpenAI notes that, for governance, the document is a model card rather than a system card, reflecting the reality that gpt-oss models will be used across a wide array of systems developed by many stakeholders. The organization highlights the necessity for stakeholders to implement their own safety decisions and safeguards while the models themselves are designed to follow OpenAI’s safety policies by default. Regarding performance, scalable capability evaluations on gpt-oss-120b indicate no High capability in the Tracked Categories of Biological/Chemical, Cyber, or AI Self-Improvement. OpenAI

Why it matters (impact for developers/enterprises)

For developers and enterprises, the release of open-weight models under Apache 2.0 provides a flexible option for integrating advanced reasoning within custom systems while maintaining a clear safety posture. The ability to adapt reasoning effort, use tools like web search and Python execution, and rely on full CoT and Structured Outputs can support a range of agentic workflows and complex decision tasks. However, this openness carries a distinct risk profile: once released, models may be fine-tuned in ways that circumvent safety features or optimize for harmful outcomes. As a result, OpenAI calls for additional safeguards by organizations seeking to replicate the protections that are typically built into API-based deployments. Stakeholders should consider their own governance, risk, and compliance frameworks before deployment. OpenAI The company also frames the gpt-oss models as part of a broader commitment to beneficial AI and to raising safety standards across the ecosystem. The model card approach acknowledges that safety architecture and protective measures may differ across systems and require ongoing oversight by the institutions implementing these models. This release is positioned within a broader effort to balance openness with safety in real-world deployments. OpenAI

Technical details or Implementation

  • Model type: Text-only, open-weight reasoning models. OpenAI
  • License: Apache 2.0, plus OpenAI’s gpt-oss usage policy. OpenAI
  • API compatibility: Designed to be used with OpenAI’s Responses API. OpenAI
  • Use cases: Agentic workflows with strong instruction following, tool use (web search, Python code execution), and adjustable reasoning effort. OpenAI
  • Capabilities: Full chain-of-thought (CoT) and Structured Outputs; customizable behavior. OpenAI
  • Safety posture: Safety is foundational; open models carry a different risk profile; safeguards may be needed to replicate system protections. OpenAI
  • Evaluation: Scalable capability evaluations on gpt-oss-120b show the default model does not meet High capability thresholds in Biological/Chemical, Cyber, or AI Self-Improvement categories. OpenAI
  • Dates: Publication Aug 22, 2025; Safety Aug 7, 2025; Publication Aug 7, 2025. OpenAI | Feature | gpt-oss-120b | gpt-oss-20b |---|---|---| | License | Apache 2.0 | Apache 2.0 |Output type | Text-only | Text-only |API compatibility | Responses API | Responses API |Key capabilities | Instruction following, tool use, adjustable reasoning, CoT, Structured Outputs | Instruction following, tool use, adjustable reasoning, CoT, Structured Outputs |Safety posture | Founded on safety policies; open models differ from proprietary models | Founded on safety policies; open models differ from proprietary models |Evaluation outcome | Did not reach High capability in tracked categories on gpt-oss-120b | (Not specified in this excerpt) |

Key takeaways

  • Open-weight models gpt-oss-120b and gpt-oss-20b are released under Apache 2.0 with a defined usage policy. OpenAI
  • They are text-only and designed for integration into complex, tool-enabled workflows. OpenAI
  • Full CoT and Structured Outputs are supported, with safety as a core consideration. OpenAI
  • The security and safety posture of open models requires additional safeguards by implementing entities. OpenAI
  • Early evaluations indicate the default gpt-oss-120b model does not reach High capability in key risk areas. OpenAI

FAQ

  • What are gpt-oss-120b and gpt-oss-20b?

    They are open-weight reasoning models released under the Apache 2.0 license and the gpt-oss usage policy, designed for text-only tasks and compatible with the Responses API.

  • What is meant by open-weight?

    It refers to models released with weights that may be used and modified under the Apache 2.0 license in open environments.

  • What safety considerations are highlighted?

    Safety is foundational; open models carry a different risk profile, and attackers could potentially fine-tune them to bypass safety or optimize for harm without additional mitigations from OpenAI. Stakeholders may need extra safeguards.

  • How were the models evaluated?

    OpenAI conducted scalable capability evaluations on gpt-oss-120b and reported that the default model did not reach High capability thresholds in Biological/Chemical, Cyber, or AI Self-Improvement categories.

  • Where can I learn more?

    See the model card linked in the references.

References

More news