Did the OpenAI onsite (virtual) in January. Five sessions across two days. Here's the unfiltered version.
Structure:
Day 1: Coding round 1 (45 min, one problem, live in CoderPad) System design (60 min, one design problem, Excalidraw) Cross-functional / PM-adjacent round (45 min)
Day 2: Coding round 2 (45 min) Behavioral / values (45 min)
Breaks between sessions were short. By session 4 I was mentally running on fumes. This is a real consideration: the energy management across a 2-day virtual onsite is different from a 1-day in-person loop.
Coding rounds: Both were 'medium that escalates.' First one I got a tree problem that started as a traversal and ended with me implementing a custom comparator for a priority queue. Second was graph-based shortest path with a dynamic edge-weight constraint they added 25 minutes in. Both times the interviewer told me upfront: 'we're going to add complexity, that's expected, we want to see how you adapt.'
System design: I wrote about this separately but briefly: design a distributed rate limiter. Deep on the storage layer (Redis vs. a custom token bucket, why, tradeoffs), deep on how you handle edge cases at the API gateway layer. 60 minutes goes fast.
Cross-functional round: A PM asked me about shipping a feature with incomplete requirements. 'Here's a hypothetical: your team ships an experimental feature but usage data is ambiguous. How do you decide whether to invest further or cut scope?' This is not a coding problem. They wanted engineering judgment + product thinking. No right answer, but they noticed if your framework was incoherent.
Behavioral round: About 4 questions, STAR format, heavy on ambiguity and judgment calls. Not 'tell me about a conflict' generic stuff. See my other posts for specifics.
Debrief timing: Got a call from the recruiter 8 days later. She said debrief went well and asked about offer preferences (equity vs. cash balance). Verbal offer in the same call. Written offer 4 business days after.
Total loop to offer: about 7 weeks. Felt long while I was in it, reasonable in retrospect.