Benchmarking biochemical networks generated by large language models