OpenAI Launches Free, Customizable AI Models to Rival Meta

OpenAI has announced the release of its own freely available artificial intelligence models, marking a strategic shift to compete with Meta and DeepSeek in the rapidly evolving AI landscape. The developer behind ChatGPT is making two “open-weight” large language models, gpt-oss-120b and gpt-oss-20b-two, available for free download and customization by users. This initiative contrasts with OpenAI’s flagship ChatGPT, which operates on a “closed” model that cannot be altered by external users.

The move by OpenAI mirrors Meta’s approach with its Llama models, which are also offered on a similar basis. Sam Altman, OpenAI’s chief executive, expressed enthusiasm for contributing to the growing collection of freely accessible AI models, emphasizing their foundation in “democratic values” and potential for “wide benefit.” He stated that making these models, the product of billions of dollars in research, available to the world aims to democratize access to AI.

This sentiment is echoed by Mark Zuckerberg, Meta’s CEO, who has similarly argued that making models freely available and customizable broadens access to AI’s benefits and opportunities, preventing power from being concentrated among a few companies. However, Meta has also indicated that it may need to exercise caution when applying this approach to highly advanced AI models.

Adding to the competitive landscape, DeepSeek, a Chinese rival to both OpenAI and Meta, has also released powerful models that can be downloaded and customized without cost.

OpenAI claims its new gpt-oss models demonstrate superior performance in reasoning tasks compared to similarly sized models. Notably, the larger gpt-oss-120b model reportedly achieved near-equal core reasoning performance to OpenAI’s o4-mini model. In terms of safety, OpenAI stated that during testing, it created “maliciously fine-tuned” versions of the models to simulate biological and cybersecurity threats, but found these versions “unable to reach high capability levels” for such harms.

The emergence of powerful, freely available, and customizable AI models has raised concerns among some experts. They warn that such models could potentially be adapted to cause serious harm, including the development of bioweapons, if misused.

The terminology surrounding model accessibility varies. Meta defines its Llama models as “open source,” implying that their training datasets, architecture, and training code are also provided alongside the free download and customization options. However, the Open Source Initiative, a US-based industry organization, argues that restrictions placed by Meta on the use of its models prevent them from qualifying as fully open source. OpenAI, in contrast, uses the term “open weight,” which signifies that while the model can be fine-tuned, its internal workings are not fully transparent to the developer—a step down from full open source.

OpenAI’s announcement comes amid ongoing speculation about the potential release of GPT-5, a new version of the foundational model for ChatGPT. OpenAI CEO Sam Altman recently shared a screenshot that appeared to be from the company’s latest AI model.

Meanwhile, Google has also outlined its latest advancements towards artificial general intelligence (AGI) with a new “world model” called Genie 3. Developed by Google DeepMind, this model allows AI systems to interact within convincing simulations of the real world. Google suggests that Genie 3 could be instrumental in training robots and autonomous vehicles by allowing them to engage with realistic recreations of environments such as warehouses. Google DeepMind posits that such “world models” are a crucial step towards achieving AGI—a hypothetical level of AI where a system can perform most tasks at a human level, rather than just specialized functions, potentially even taking on human jobs. These models are expected to play a significant role in the development of AI agents, which are systems designed to carry out tasks autonomously.