The 12 coding questions OpenAI keeps asking. Its screen bans autocomplete.

• 12 coding questions are reported at OpenAI more than any others, and 10 of them were asked again since Sep 1, 2026. A GPU credit ledger, a memory allocator, a follow graph with snapshots: work you could ship, built in parts.
• The coding screen allows no AI and no autocomplete. Documentation is fine. The onsite coding round allows no AI either.
• 4 of the 12 are ML coding in PyTorch or NumPy, though ML coding is only 8 of all 84 coding questions.
• 31 of 50 design questions are classic system design. A chess site, a payment system, a sandboxed cloud IDE.
• The recruiter call is a real gate. A weak answer on AI safety can end the process there.

OpenAI’s loop has 7 rounds, and both coding rounds rule out AI

The loop starts with a 30-minute recruiter call. Then come two 60-minute technical screens, one in coding and one in system design; for some roles they fall on the same day. The final loop is virtual: onsite coding, a second design round, a project presentation and a behavioral round, then team matching. Ask the recruiter what your loop contains, because it changes by team. Each round and its length:

1. Recruiter call30 min
2. Technical screen: coding60 min · no AI, no autocomplete
3. Technical screen: system design60 min
4. Onsite coding60 min · no AI
5. Onsite system design60 min
6. Project presentation45 min
7. Behaviorallength varies

The coding screen rules out AI assistants and autocomplete; you may read documentation. It is one problem in three to five growing parts, usually in CoderPad. ML roles get a Colab notebook with autocomplete off. Onsite coding allows no AI and runs 60 minutes, on a problem larger than the screen’s. Some infra loops add a 75-minute systems coding round on concurrency. If you code with an AI assistant every day, turn it off for your practice weeks.

The technical screen produces more reported questions than any other round

We sorted OpenAI’s 144 reported questions by the round candidates named (n=136 with a known round). The screen leads with 45, ahead of the coding round (29) and the online assessment (18). Some early-career candidates take a timed online test before anything else.

1. Technical screen45 questions
2. Coding round29 questions
3. Online assessment18 questions
4. System design17 questions
5. AI system design8 questions
6. Onsite coding7 questions
7. Recruiter call4 questions
8. ML design4 questions

41 of 84 coding questions were reported once. 12 keep coming back

OpenAI’s coding questions by how much candidates report them (n=84). You can’t memorise a long tail like that. The top tier is short enough to build every one:

1. Reported the most12 questions
2. Medium10 questions
3. Low16 questions
4. Reported once41 questions

The 12 reported the most, newest first. Each was last asked between Jul 10, 2026 and Oct 13, 2026:

• Find noisy annotators and retrain a classifier without them (technical screen), last asked Oct 13, 2026
• Memory allocator with malloc, free and coalescing (technical screen), last asked Oct 8, 2026
• Threshold infection spread on a grid (technical screen), last asked Oct 8, 2026
• Find the planted bugs in a small Transformer training script (onsite coding), last asked Oct 2, 2026
• GPU credit ledger with out-of-order events (coding round), last asked Sep 29, 2026
• Matrix product scan with autograd and a hand-written backward (coding round), last asked Sep 27, 2026
• Vectorized nearest-neighbour classifier as a linear layer (coding round), last asked Sep 26, 2026
• Social follow graph with point-in-time snapshots (coding round), last asked Sep 20, 2026
• Toy language type inference with generics and tuples (technical screen), last asked Sep 20, 2026
• Count and map a cluster tree with message passing (coding round), last asked Sep 10, 2026
• Turn-based monster team battle simulator (coding round), last asked Aug 18, 2026
• Data labeling scheduler with prefix-balanced models (coding round), last asked Jul 10, 2026

One tier down, 6 more were asked in the same weeks:

• Fault-tolerant work queue with leases and retries (coding round), last asked Oct 16, 2026
• Persistent key-value store with binary snapshots (coding round), last asked Oct 13, 2026
• Extensible chat app with command bots (coding round), last asked Oct 7, 2026
• Three-card showdown between two players (technical screen), last asked Oct 2, 2026
• IPv4 range iterator with reverse and CIDR support (coding round), last asked Sep 23, 2026
• Entropy from logits, stable and then streaming (technical screen), last asked Jul 26, 2026

Read the titles and you see work, not puzzles: a ledger that takes events out of order, an allocator that coalesces free blocks, type inference for a toy language. In the screen, two or three parts done cleanly is the usual bar, and a finished, tested solution beats an unfinished clever one.

4 of the top 12 are ML coding, against 8 of all 84

ML coding is a small share of OpenAI’s coding questions, yet 4 of the 8 sit in the top tier: find noisy annotators and retrain a classifier without them, find the planted bugs in a small Transformer training script, matrix product scan with autograd and a hand-written backward and vectorized nearest-neighbour classifier as a linear layer. ML and research roles get questions like these in place of the usual problems. You write PyTorch or NumPy by hand, with no autocomplete to fill in a tensor shape, so practise until broadcasting and a backward pass come without looking anything up.

31 of 50 design questions are classic system design

OpenAI’s design questions by type (n=50). The screen and the onsite each have a 60-minute design round. The screen grades a concrete API and storage schema; the onsite grades scaling under fast growth and correctness in production: retries and idempotency. The design questions reported the most: design a sandboxed cloud IDE, design an online chess game, design a payment system, design a distributed power grid control system and design a video generation service on a scarce GPU pool. ChatGPT-like systems come up too, and you may be told to treat model serving as a black box. Our post on AI lab design questions compares OpenAI’s mix with the other labs’.

1. Classic system design31 questions
2. AI system design12 questions
3. ML design7 questions

The recruiter call can end your loop over AI safety

The recruiter call asks for your background, why OpenAI, a recent project, and your view on AGI and AI safety. A weak answer on AI safety can end the process right there. The behavioral round comes last, only after the technical rounds pass, and the hiring manager can be the one running it. Expect to be asked what you would do if told to ship without safety testing. How you do there weighs on the level you get. The conversation questions reported the most are recruiter screen: motivation, background and logistics, project deep dive: present a project you led and hiring-manager round: why OpenAI, AGI and safety dilemmas.

The project presentation runs 45 minutes: a few slides on one or two projects you led, then rapid questions on what you owned, why you chose it over the alternative and what happens at ten times the scale. Candidates report strong coders failing it by staying shallow. Our deep dive post covers how the AI labs run this round.

What this means for your prep

1. Build the 12 coding questions first, with autocomplete off. 10 of them were asked again since Sep 1, 2026. Do each one in parts, and write tests as you go.
2. Add the 6 from the next tier once the top tier is done. Skip the long tail; 41 of 84 coding questions were reported only once.
3. For an ML or research role, drill the 4 ML coding questions in plain PyTorch and NumPy.
4. Practise classic design. 31 of 50 design questions are classic system design. Bring a concrete API and schema, and say how retries stay idempotent.
5. Write down your view on AI safety before the recruiter call, and pick the one or two projects you will present. Both are gates.