While we have been building in the background, we have been running benchmarks SERV Reasoning outperforms GPT-5 on accuracy, costs 96% less, and produces consistent results across runs The public API is not live yet. This is just what the early numbers look like