OPEN SOURCE AI JUST BEAT OPENAI’S BEST SEARCH MODEL AND THE RESEARCHERS ARE GIVING IT AWAY FOR FREE A team of researchers from the University of Illinois Urbana-Champaign, UC Berkeley and the open-source vector database platform Chroma has released Harness-1, a 20-billion parameter AI search agent trained on a fraction of the data competing models required, and it outperforms OpenAI’s GPT-5.4 on the task it was built for. The benchmark that matters here is recall: the ability to surface the right information from a large document set when a user or agent needs it. Harness-1 achieved a 73 percent average performance score on this task, compared to GPT-5.4 at 70.9 percent. It also outperformed Anthropic’s Sonnet 4.6 and several other leading models on the same evaluations. Only Opus 4.6 consistently edged it out. The technical innovation behind Harness-1 is an architecture the researchers call a harness, which externalizes working memory from the execution environment. Instead of forcing the model to hold candidate documents, evidence and links inside its own context window, the harness manages that state externally. The model’s cognitive load drops. Its retrieval accuracy goes up. The entire model was trained on roughly 4,400 unique items. Competing open-source models required between 17,000 and 221,000. That efficiency gap matters as much as the benchmark number. It means smaller labs and independent researchers can now build competitive search agents without industrial-scale compute budgets. The open-source release puts serious pressure on the paid search API market. Keywords: Harness-1, open source AI, AI search agent, GPT-5.4 alternative