OpenAI gpt-oss-120b and gpt-oss-20b Release

OpenAI has released gpt-oss-120b and gpt-oss-20b, two open-weight reasoning models designed for agentic workflows. These models are available under the Apache 2.0 license and the gpt-oss usage policy, providing the open-source community with tools for strong instruction following, tool use, and customizable reasoning effort.

Model Capabilities and Integration

The gpt-oss models are text-only reasoning models that support Structured Outputs and provide full chain-of-thought (CoT) visibility. They are designed for integration into agentic workflows, specifically supporting tool use such as Python code execution and web search.

Key technical features include:

  • Customizability: The models are customizable to specific needs.
  • Instruction Following: Strong capabilities in following complex instructions.
  • Reasoning Effort: Users can adjust the reasoning effort for tasks that that do not require complex reasoning.
  • API Compatibility: The models are compatible with the OpenAI Responses API.

Safety and Risk Assessment

Because open-weight models can be fine-tuned by third parties to bypass safety refusals, OpenAI treats them as having a different risk profile than proprietary models. OpenAI notes that developers and enterprises implementing these models will need to implement their own system-level protections to replicate the protections found in OpenAI's proprietary API products.

To evaluate the risk of these models, OpenAI conducted scalable capability evaluations on gpt-oss-120b using its Preparedness Framework. The evaluations focused on three Tracked Categories: Biological and Chemical capability, Cyber capability, and AI Self-Improvement. The results confirmed that the default model does not reach the threshold for "High capability" in any of these categories.

Adversarial Testing and Open-Model Frontier

OpenAI conducted adversarial fine-tuning simulations to determine if gpt-oss-120b could be modified by an attacker to reach High capability in the Biological and Chemical or Cyber domains.

According to the findings:

  • Adversarial Fine-tuning: Even with robust fine-tuning using OpenAI's training stack, gpt-oss-120b did not reach High capability in Biological and Chemical Risk or Cyber risk.
  • Comparison to Existing Models: The release of gpt-oss-120b does not significantly advance the frontier of biological capabilities in open foundation models, as existing open models' default performance often matches the performance of the adversarially fine-tuned gpt-oss-120b.

OpenAI states that this release is part of a commitment to advancing beneficial AI and raising safety standards across the AI ecosystem.

Sources