OpenAI has released two new open-weight language models, gpt-oss-120b and gpt-oss-20b, marking its first open-source model release since GPT-2 in 2019. These models are available under the Apache 2.0 license, allowing free use, modification, and redistribution, including commercial applications. The gpt-oss-120b model features approximately 117 billion parameters, uses a Mixture-of-Experts architecture, and offers performance comparable to OpenAI's o4-mini model on reasoning benchmarks. It can run on a single Nvidia H100 80GB GPU and supports advanced features such as a 128,000 token context window, chain-of-thought reasoning, and agentic capabilities. The smaller gpt-oss-20b model has about 21 billion parameters, matches the o3-mini model's performance, and can operate on consumer-grade hardware with as little as 16GB of GPU memory, enabling deployment on laptops and edge devices. Both models are text-only and designed for reasoning-heavy tasks, including code, science, and math. OpenAI developed these models with community feedback to enhance reasoning capabilities and safety. Amazon Web Services (AWS) has integrated these models into its cloud offerings, making them available via Amazon Bedrock and Amazon SageMaker for the first time, providing AWS customers access to OpenAI's open-weight models. AWS highlighted that the larger gpt-oss-120b model delivers three times better price-performance compared to Google's Gemini model and five times better than DeepSeek-R1. This release represents a strategic shift for OpenAI, returning to its open-source roots and expanding accessibility to powerful AI models across a broad range of users and applications.