flow-1 is billed as a cheaper way to find errors in AI agent traces

The creator of flow-1 says the model was trained with reinforcement learning to find errors in AI agent traces. They claim it matches GPT-6-sol in trace intelligence while being 23x cheaper, costs 25% less to run than GPT-6-luna, and makes it possible to monitor and understand every agent run without sampling.