The Great Open Source Lie: How OpenAI's gpt-oss-120b Isn't Open Anything You Can Actually Use
SAN FRANCISCO — In a press conference so heavily lawyered that the lawyers themselves needed to consult a legal AI to ensure they were speaking in proper third-person passive voice, OpenAI announced today it was “returning to its roots” by releasing two new large language models under the gpt-oss designation: the behemoth 120-billion-parameter gpt-oss-120b and the allegedly “lightweight” gpt-oss-20b. The company framed this as a mission to democratize access to “open source AI,” a phrase that would be a major understatement if these models could actually be used to build something other than a compliance dashboard.