Stay informed with free updates
Simply sign up to the Artificial intelligence myFT Digest — delivered directly to your inbox.
OpenAI has released its first open artificial intelligence model since it launched ChatGPT, as the $300bn group attempts to quash rising competition from Chinese start-up DeepSeek and others in the cutting-edge technology.
The San Francisco-based company unveiled two “open-weight” models on Tuesday that will be free to access and for developers to customise, providing a more transparent alternative to its existing closed AI offerings.
Their launch comes six months after advances by China’s DeepSeek caused shockwaves following the release of its open model R1 in January, whose performance was comparable to some of OpenAI’s products, undermining Silicon Valley’s lead in a global AI arms race.
Days after R1 was unveiled, OpenAI chief executive Sam Altman said he believed his company had “been on the wrong side of history here and need[ed] to figure out a different open-source strategy”.
OpenAI’s new models, called “gpt-oss”, perform as well as some of its smaller closed models that power ChatGPT, and were “designed to be used within agentic workflows”, meaning the system operates autonomously, and has been trained to process complex queries step by step. Developers will also be able to adjust the amount of effort the model puts into this “reasoning”, rather than responding quickly.
While the new OpenAI models are “open-weight” rather than “open-source” — which provides more comprehensive information, including datasets and the code required to train a model from scratch — they had comparable performance to leading open models, including from China. OpenAI did not provide details on how it might monetise the open product.
“OpenAI’s mission is to ensure [artificial general intelligence] that benefits all of humanity,” said Altman on Tuesday. “To that end, we are excited for the world to be building on an open AI stack created in the United States, based on democratic values, available for free to all and for wide benefit.”
The success of DeepSeek and other Chinese AI systems, including Alibaba’s Qwen and Moonshot’s Kimi, has helped China to take the lead from the US in open-source AI technology and spurred huge demand from developers. In the US, rival Meta has focused on developing open-weight models, but its recent system has fallen short of expectations.
OpenAI’s larger oss model performed similarly to its closed o4-mini model, while its smaller version had results close to its o3-mini system. The small model required less memory to run, “making it ideal for on-device use cases”, such as phones or laptops, the company said.
Open-weight models are often seen as higher risk than closed systems, due to their ability to be customised and because they cannot be easily recalled if they are flawed.
OpenAI has delayed the release of the models twice, after they were initially slated for June. In July, Altman said the company needed time to “run additional safety tests”.
Recommended
On Tuesday it detailed how it had created a version of the models to deliberately mimic how a bad actor might use the new products. They were then tested for vulnerabilities that could allow the AI to be misused, such as designing biological weapons or creating novel viruses.
The malicious models were “unable to reach high capability levels” in internal tests and were reviewed by three independent expert groups, who recommended improvements for testing, it added.
Meta’s chief executive Mark Zuckerberg has long been a vocal advocate for the open AI ecosystem. However, in a memo last week, he said Meta’s new push to develop “superintelligent systems” that surpass the intelligence of humans “will raise novel safety concerns”, adding: “We’ll need to be rigorous about mitigating these risks and careful about what we choose to open-source.”
Additional reporting by Hannah Murphy in San Francisco and Eleanor Olcott in Beijing