OpenAI onsite coding interview
Round 4 of 7 · 29 questions reported
- Length
- 60 min
- Format
- Live coding; the problem often rebuilds a familiar system in several parts
- They grade
- Correct code first, then performance; test discipline; command of your language (iterators, generators, async in Python)
- AI use
- No AI
- Heads-up
- Larger in scope than the screen, and speed matters. Infra loops add a separate ML coding round on top of the systems coding round.
Questions OpenAI asks in this round 29
- GPU credit ledger with out-of-order eventsSep 2026 · very often
- Social follow graph with point-in-time snapshotsAug 2026 · very often
- Find the planted bugs in a small Transformer training scriptOct 2026 · often
- Vectorized nearest-neighbour classifier as a linear layerSep 2026 · often
- Count and map a cluster tree with message passingSep 2026 · often
- Matrix product scan with autograd and a hand-written backwardAug 2026 · often
- Turn-based monster team battle simulatorAug 2026 · often
- Data labeling scheduler with prefix-balanced modelsJul 2026 · often
- Extensible chat app with command botsOct 2026 · sometimes
- Fault-tolerant work queue with leases and retriesOct 2026 · sometimes
- IPv4 range iterator with reverse and CIDR supportSep 2026 · sometimes
- Persistent key-value store with binary snapshotsAug 2026 · sometimes
- Typed task lock for mutually exclusive job typesAug 2026 · rarely
- Sharded matrix multiply: forward, backward and planted bugsJul 2026 · rarely
- Resumable iterators that save and restore their positionJun 2026 · rarely
- Read and refactor a large PyTorch training fileMar 2026 · rarely
- In-memory SQL table with filtering and multi-key sortingNov 2025 · rarely
- Temporal key-value store with point-in-time reads and restoreOct 2025 · rarely
- Resolve a cd command to an absolute pathDec 2025 · once
- Mini spreadsheet with formulas and cycle detectionNov 2025 · once
- Alert routing service with silences and escalation
- Debug a Python job scheduler
- Earliest release that supports a feature
- Guess the number with one-call delayed feedback
- Indexed sequence with checkpoints
- Per-user chat event counts in a sliding window
- Rectangle board with remove-at-cell, batch remove and undo
- Shuffle, deal and sort a four-player card hand
- Work queue with a dead-letter queue and dependency cycle detection
About the OpenAI onsite coding round
What does OpenAI ask in the onsite coding round?
The most asked: GPU credit ledger with out-of-order events, Social follow graph with point-in-time snapshots and Find the planted bugs in a small Transformer training script. 29 questions reported in all.
How long is the OpenAI onsite coding round?
About 60 minutes.
What is the OpenAI onsite coding round like?
Live coding; the problem often rebuilds a familiar system in several parts. They grade: Correct code first, then performance; test discipline; command of your language (iterators, generators, async in Python).
Can you use AI in the OpenAI onsite coding round?
No AI.