OpenAI announced the release of GPT‑OSS, comprising two open‑weight language models—gpt‑oss‑120b and gpt‑oss‑20b—designed for high-performance reasoning and autonomous agentic workflows. According to the official OpenAI release, these are the company’s first “open‑weight” models since GPT‑2 back in 2019, making their trained weights freely accessible for developers to run and fine‑tune on their own systems. Unlik... The gpt‑oss‑120b variant reportedly matches or exceeds the performance of OpenAI’s o4‑mini model on core reasoning tasks, while gpt‑oss‑20b is optimized to run on a single GPU and standard laptops. Both excel in coding, mathematics, and health-related reasoning benchmarks. OpenAI emphasized that the models were rigorously evaluated using a “Preparedness Framework,” including tests involving malicious fine-tuning. According to the company, the worst-case versions “did not reach high capability levels,” demonstrating limited unsafe behavior. OpenAI also collaborated with external experts on safety advisory measures. Amazon Web Services has already added these models to its Amazon Bedrock and SageMaker JumpStart platforms, marking the first time OpenAI’s models are offered directly through AWS cloud services. The gpt‑oss‑120b model is touted to be up to three times more cost-efficient than similar models offered by Gemin, DeepSeek‑R1, and o4-mini. CEO Sam Altman described GPT‑OSS as built “based on democratic values … to get AI into the hands of the most people possible.” Despite not providing full transparency like open-source releases, the open‑weight approach is being positioned as a strategic bridge between accessibility and robustness.