They released some examples of what their workflow evals are like. I'm sure you could reverse engineer a benchmark from that

https://evals.typesafe.ai/