This article discusses Meta's TestGen-LLM tool, which leverages large language models (LLMs) to automatically enhance existing manually written tests. TestGen-LLM ensures that its generated test classes successfully pass a series of filters, guaranteeing measurable improvements over the original test suites and mitigating issues caused by LLM hallucinations. We detail the deployment of TestGen-LLM during Meta's testing marathons on the Instagram and Facebook platforms. When evaluating the tool on Instagram's Reels and Stories products, 75% of TestGen-LLM's test cases were constructed correctly, 57% reliably passed, and 25% expanded the coverage of the tests. During the testing marathons on both Instagram and Facebook, TestGen-LLM improved 11.5% of the classes it was applied to. Notably, 73% of these suggestions were accepted by Meta's software engineers and integrated into production deployments. This marks the first reported large-scale industrial deployment of LLM-generated code with such a high degree of success in improving code quality. The results demonstrate the potential of LLMs in automating and enhancing the software testing process, offering significant benefits for developers and organizations.