OpenAI Optimizes ChatGPT And Unveils Two Open-Weight gpt-oss Reasoning Models

After years of proprietary AI models with limited access, the generative AI landscape is beginning to evolve. As demand for more transparency and control grows, especially from researchers and enterprises, the focus has increasingly turned to open-weight and open-source models.

OpenAI, long known for its guarded approach to AI innovation, is now entering the growing field of open-weight models.

On Tuesday, OpenAI announced two open-weight language models: gpt-oss-120b and gpt-oss-20b. These models are built for advanced reasoning and optimized to run locally, including on laptops.


Trained on a text-only dataset rich in math, science, and coding knowledge, the models match the performance of OpenAI’s proprietary o3-mini and o4-mini models, the company said.

"One of the things that is unique about open models is that people can run them locally. People can run them behind their own firewall, on their own infrastructure," said Greg Brockman, co-founder of OpenAI, during a press briefing.

The release marks OpenAI’s first open model launch since GPT-2 in 2019. Unlike open-source models, open-weight models offer public access to trained parameters, allowing developers to fine-tune them without original training data.

In a notable move, Amazon Web Services (AWS) has also made these models available on Amazon Bedrock, making them the first OpenAI models on the platform. “OpenAI has been developing great models and we believe that these models are going to be great open-source options, or open-weight model options for customers,” said Atul Deo, director of product at Bedrock.


In a parallel effort, OpenAI also published a study evaluating the worst-case frontier risks of releasing gpt-oss. The paper introduces malicious fine-tuning (MFT), a stress test to gauge how dangerous the open-weight model could be if deliberately optimized for harmful activities.

The team trained gpt-oss in two sensitive areas: biology, using web browsing in a reinforcement learning setup to simulate threat creation; and cybersecurity, where it was tested in solving capture-the-flag (CTF) challenges within a coding environment.

Even after aggressive fine-tuning, gpt-oss underperformed OpenAI’s own o3 model in both domains. While it showed minor increases in biological risk compared to other open-weight models, it did not push the current frontier of threat potential.

These results, OpenAI said, were key to its decision to release the models, highlighting the balance between capability and responsible openness.

Following the release of its open-weight models, OpenAI is continuing to shed light on its philosophy behind the design and purpose of ChatGPT.

In a recent update, the AI leader outlined its intention to optimize ChatGPT to help users make progress, learn something new, or solve a problem, and then return to their lives. “Our goal isn’t to hold your attention, but to help you use it well,” the post states.

Instead of measuring success through clicks or time spent, OpenAI looks at whether users leave having accomplished what they came for, and whether they find the product helpful enough to return.

This philosophy extends to how ChatGPT supports users during more personal or emotionally complex situations. OpenAI acknowledged that an earlier update made the model too agreeable, at times saying what sounded comforting rather than what was genuinely helpful.

The company has since rolled it back, adjusted its feedback processes, and is now working toward long-term improvements in real-world usefulness.


Key changes include training ChatGPT to recognize signs of mental or emotional distress more effectively, particularly where users may be vulnerable. The system is also being designed to offer grounded honesty and guide decision-making without taking control, especially in high-stakes personal matters. “ChatGPT shouldn’t give you an answer. It should help you think it through,” OpenAI noted.

To support these changes, OpenAI has consulted over 90 physicians across 30 countries, partnered with human-computer-interaction (HCI) researchers and clinicians, and is forming an advisory group of experts in mental health and youth development.

The effort is grounded in one test: “If someone we love turned to ChatGPT for support, would we feel reassured?” The company says reaching an unequivocal "yes" remains its benchmark.

This move by OpenAI highlights how open-weight AI models can gain the public’s trust through responsible research and deployment. Plus, its upgrades to optimize ChatGPT will make the AI experience better for everyone.

Do you think AI companies should involve mental health experts to make them safer and more supportive? Do the benefits of open-weight AI models outweigh their potential risks?

Let us know your thoughts in the comments section below!