Company interview guide

OpenAI Machine Learning Interview Guide (2026)

Practical coding, hands-on ML debugging and LLM system design, with a strong mission check.

OpenAI's machine learning interview loop has 7 reported stages, from recruiter call to behavioral, deep dive & hiring manager; typical length: 4-10 weeks.

Roles covered: ML Engineer, Research Scientist and AI Engineer. The question bank holds 95 questions reported from OpenAI interviews, 16 of them free to solve in your browser.

Try a free question

Interview process length
4-10 weeks
Stages
7
Reports linked to bank questions
117
Bank report dates
Nov 2024–Sep 2026
Questions in bank
95 (16 free)
Last reviewed
October 2026

How the OpenAI interview loop works

A resume review, a recruiter or hiring-manager call, a technical screen of usually two back-to-back 60-minute rounds, then the main loop. For ML and applied engineering roles the main loop is four to six interviews over one or two days: coding, ML coding, ML / LLM design, behavioral and a deep dive on your own work. Research-track candidates in 2026 often describe about four coding rounds (two general, two ML), sometimes split across two sittings a week apart, before a final stage with the hiring manager and a research presentation; some add a math and reasoning round. Team placement is mostly settled early, and the company says a decision usually follows within about a week. Take-home projects are rare for ML roles in our reports. Rescheduling is common, so build slack into your timeline.

The OpenAI interview process, stage by stage

01Recruiter call

About 30 min with a recruiter or the hiring manager

A conversation about your background, what draws you to OpenAI and which teams interest you. Several candidates say team preference was settled here, and a few had a short hiring-manager screen with the target team before any coding, so come with a clear view of where you want to work.

What it tests

  • Why OpenAI, and why now
  • How your past work maps to the team you are targeting
  • Logistics: timeline, location, level

How to prepare

  • Ask for the type, length and tooling of every later round; the answer predicts the content.
  • Some candidates get a written prompt per round and others only a generic preparation document, so ask if nothing arrives.
  • Have a two-minute summary of your most relevant project ready.

Recruiter or hiring-manager intro call in 11 ML and research reports since 2024 plus 2 other-role reports; team preference came up in five. The company's interview guide lists this call with a recruiter or hiring manager.

02Technical screens

Usually 2 x 60 min back to back: CoderPad for general coding, a notebook for ML coding; some research screens run three or four rounds

Two consecutive rounds with no break, drawn from the same pool as the later rounds. Reports from 2026 mostly pair a multi-part coding task with an ML coding task in numpy; 2025 reports more often paired the transformer bug hunt with a RAG design. Some research screens add a math round or a debugging round, use a 75-minute ML slot, or run in two sittings about a week apart.

What it tests

  • Writing clean, modular code for a problem whose rules grow part by part
  • Simulation with careful state updates (for example, day-by-day grid spread)
  • Turning a linear search into a faster one by spotting hidden structure
  • Vectorized numpy without loops, plus numerically stable reductions
  • Staying sharp for two hours with no break

How to prepare

  • Aim to finish three or four of the parts; interviewers say up front that finishing everything is not required.
  • Update simulation state synchronously (compute the next grid from the current one) to avoid order bugs.
  • Practise in a plain editor with autocomplete off and only the standard library.
  • Do at least one full two-hour mock to rehearse the fatigue.

Free reported questions for this stage

Described in 29 ML and research reports since 2025; 12 give two rounds, the rest one or a three-to-four-round research screen. The company's guide calls it a skills-based assessment (pair coding, take-home or tests) varying by team.

03Main loop: general coding

1-2 x 60 min in CoderPad, standard library only, with starter code and runnable tests

Build a small but realistic system, such as a key-value store, a scheduler or a ledger, and extend it as new requirements arrive; the multi-part simulations from the screen also recur here. Some rounds labelled coding turn out to be debugging an existing codebase. Research loops usually include one or two of these rounds.

What it tests

  • Data modelling and clean interfaces for a small in-memory system
  • Extending your own design as the interviewer adds rules
  • Handling ordering, versioning and expiry edge cases
  • Debugging unfamiliar code: race conditions, deadlocks, rate limiting
  • Writing tests, sometimes before the implementation

How to prepare

  • Clarity and correctness count for more than raw speed; keep functions small and named well.
  • When debugging, start from the failing test or assertion instead of reading the whole repository.
  • Design for change: assume the next part will add a rule you have not seen.

Free reported questions for this stage

General coding after the screen in 11 ML and research reports since 2025, usually one or two 60-minute rounds; one got a debugging session instead. The company's guide stresses code quality, performance and tests in finals.

04Main loop: ML coding & debugging

1-2 x 60-75 min in a notebook with numpy / PyTorch and starter code; tests are provided in some rounds but not all

The most reported ML round: repair a small transformer with four or five planted bugs, then add a KV cache. It has come up in the screen about as often as in the main loop. Other variants ask for hand-written backward passes, a numpy puzzle, or a data-centric classifier task on noisy annotator labels.

What it tests

  • Reading and fixing transformer code: causal mask, label shift, positions, initialization
  • Implementing incremental decoding with a KV cache that matches uncached output
  • Deriving and coding backward passes (matmul, prefix products, cross-entropy)
  • Improving a model by cleaning or reweighting noisy training data
  • Explaining cost and complexity of the operations you write

How to prepare

  • Be able to fix a nanoGPT-sized model quickly; check mask direction and the one-token label shift first.
  • After adding a cache, prove it works by comparing generated tokens with and without it.
  • Know why in-place updates break autograd and how a parallel scan computes prefix products.
  • A 75-minute slot tends to mean autograd or scan work; a 60-minute slot tends to mean numpy and stability.

Free reported questions for this stage

1 more free questions for this stage are in the app.

ML coding after the screen in 14 ML and research reports since 2025, consistently in a notebook; the transformer bug hunt appears in 12 reports. The company's guide says finals focus on your area of expertise, with coding judged on quality.

05Final loop: ML / LLM system design

About 60 min, in the screen or the main loop, open discussion with a shared doc or whiteboard

Most reports describe designing retrieval-augmented assistants or enterprise search, often as the second screen round in 2025. Newer prompts cover agent harnesses with evaluation, how an assistant should behave when it does not know, data-efficient labelling at very large scale, and team-specific products. Some research loops skip this round.

What it tests

  • Retrieval design: ingestion, hybrid search, reranking, access control, freshness
  • Choosing an evaluation metric for every stage of the pipeline
  • Agent and tool-use design with safe failure handling
  • Latency, cost and long-context trade-offs
  • Scoping a vague brief into a concrete system

How to prepare

  • Have a ready RAG blueprint and a separate one for agent evaluation; interviewers have steered away from classic recommender designs.
  • Time runs short in this round, so agree on scope in the first five minutes.
  • Say how you would measure each component before you go deep on any of them.

Free reported questions for this stage

ML or LLM design round in 14 ML and research reports since 2025, eight at the screen and six later; eight of them were RAG or enterprise search prompts.

06Research rounds (math, stats, presentation)

Research-engineer and research-scientist tracks: a math and reasoning round in some screens, plus a research presentation and sometimes a collaboration interview at the end

Research loops add rounds that probe probability and derivations, plus a talk on your own work where the experimental choices are challenged. One team also ran a collaboration interview on the state of the art in the candidate's area. Formats vary by team, so ask your recruiter which apply.

What it tests

  • Probability reasoning, for example about when sampling stops
  • Deriving estimators and posteriors and coding them in numpy
  • Information-theoretic quantities used in LM training
  • Defending the design and results of a past project
  • Discussing the state of the art in your area and where it is heading

How to prepare

  • Prepare one project in depth: motivation, key decisions, what failed, what you would change.
  • Practise deriving maximum-likelihood and Bayesian answers on paper before coding them.

Free reported questions for this stage

Math, probability or statistics rounds in six research-track reports and a presentation of past work in four since 2024; formats differ by team and few give details.

07Behavioral, deep dive & hiring manager

1-3 conversations of 30-60 min, usually including the hiring manager; ML engineering loops add a deep dive on your resume, research loops add team chats

Motivation and alignment with OpenAI's mission carry real weight. Expect detailed follow-up on anything you claim about past projects.

What it tests

  • A specific, honest answer to why OpenAI
  • Collaboration, disagreement and cross-team work
  • Working well under pressure and ambiguity
  • Depth behind the claims on your resume

How to prepare

  • Read OpenAI's Charter and recent posts from the team you are joining, and connect them to your own goals.
  • Pick two or three stories and rehearse the follow-ups: scale, trade-offs, your personal contribution.
  • Do not overstate your role; deep dives quickly expose it.

Behavioral, hiring-manager or deep-dive rounds in 11 ML and research reports since 2024 plus 2 other-role reports; why OpenAI and mission fit recur. The company's guide stresses mission alignment, collaboration and communication.

Questions and themes reported most often

  • Fix several planted bugs in a small transformer until it trains (12 reports, main loop: ml coding & debugging)
  • Multi-part grid infection simulation where every part adds a rule (9 reports, technical screens)
  • Vectorized 1-nearest neighbour, then rewritten as a linear layer (9 reports, technical screens)
  • Train a classifier on noisy or dirty labels and beat the baseline (8 reports, main loop: ml coding & debugging)
  • Design an enterprise RAG or search assistant with per-stage evaluation (8 reports, final loop: ml / llm system design)
  • Matrix prefix products: forward, backward, then a parallel scan (6 reports, main loop: ml coding & debugging)
  • Deep dive or presentation on a past project (6 reports, behavioral, deep dive & hiring manager)
  • Add a KV cache to a decoder and match the uncached output (5 reports, main loop: ml coding & debugging)

What OpenAI asks about

  • ML coding & debugging37%
  • General coding28%
  • ML / LLM system design18%
  • Behavioral & deep dive10%
  • ML & LLM concepts7%

OpenAI interview prep checklist

  • Fixed a buggy mini-GPT and added a KV cache whose output matches the uncached model
  • Solved a multi-part grid simulation in under an hour without autocomplete
  • Wrote matmul and prefix-product backward passes from scratch
  • Can sketch a permission-aware RAG design with a metric for every stage
  • Have a crisp, personal answer to why OpenAI
  • Asked the recruiter for the type, length and tool rules of every round
  • Prepared a deep dive on one project, including what did not work

Tips

  • Make the transformer bug hunt plus KV cache your first priority; it is the most frequently reported ML round.
  • Practise multi-part problems against the clock and stop polishing once the tests pass.
  • Write backward passes by hand in numpy until matmul, cumulative products and cross-entropy feel routine.
  • Get fluent with broadcasting: batched linear layers, pairwise distances and log-sum-exp without loops.
  • Rehearse a notebook workflow for noisy data: explore, baseline, clean or reweight, then report macro-F1 and a confusion matrix.
  • Use the round descriptions your recruiter sends to decide what to drill in the final week.
  • Check the preparation materials for tool rules: some interviews allow AI tools and others test you without them.
  • Expect fewer hints than at large tech companies, and plan a short mental reset between back-to-back rounds.

See the rest of the OpenAI loop

The later stages, the questions reported for them and how to prepare are in Pro.

Unlock with Pro

Practice the OpenAI loop

Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.

OpenAI interview FAQ

How many interview rounds does OpenAI have for ML roles?

Candidates report 7 stages: Recruiter call, Technical screens, Main loop: general coding, Main loop: ML coding & debugging, Final loop: ML / LLM system design, Research rounds (math, stats, presentation) and Behavioral, deep dive & hiring manager. A resume review, a recruiter or hiring-manager call, a technical screen of usually two back-to-back 60-minute rounds, then the main loop. For ML and applied engineering roles the main loop is four to six interviews over one or two days: coding, ML coding, ML / LLM design, behavioral and a deep dive on your own work. Research-track candidates in 2026 often describe about four coding rounds (two general, two ML), sometimes split across two sittings a week apart, before a final stage with the hiring manager and a research presentation; some add a math and reasoning round. Team placement is mostly settled early, and the company says a decision usually follows within about a week. Take-home projects are rare for ML roles in our reports. Rescheduling is common, so build slack into your timeline.

How long does the OpenAI ML interview process take?

Reported timelines run 4-10 weeks.

What does OpenAI ask in ML interviews?

By reported question mix: ML coding & debugging 37%; General coding 28%; ML / LLM system design 18%; Behavioral & deep dive 10%; ML & LLM concepts 7%. The most repeated themes: Fix several planted bugs in a small transformer until it trains; Multi-part grid infection simulation where every part adds a rule; Vectorized 1-nearest neighbour, then rewritten as a linear layer.

How should I prepare for the OpenAI ML interview?

Make the transformer bug hunt plus KV cache your first priority; it is the most frequently reported ML round. Practise multi-part problems against the clock and stop polishing once the tests pass. Write backward passes by hand in numpy until matmul, cumulative products and cross-entropy feel routine.

Are there free OpenAI interview questions?

Yes. LeetML has 95 questions reported from OpenAI interviews, and 16 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources