DEV Community: Land the Lab

72% of Anthropic's design questions are AI or ML. 62% of OpenAI's are classic.

Land the Lab — Tue, 20 Oct 2026 00:06:29 +0000

Photo: Lucas Kepner on Unsplash

21 of Anthropic's 29 reported design questions are about AI systems or ML, against 19 of OpenAI's 50. The one reported the most is a batched LLM inference API over GPUs, last asked Oct 15, 2026.

21 of Anthropic’s 29 design questions are about AI systems or ML. At OpenAI it is 19 of 50.
The one reported the most is a batched LLM inference API over GPUs, last asked Oct 15, 2026. 3 of the 5 reported the most are AI infrastructure.
The round runs 55 minutes, with no AI, and you may type the design into a shared document. It grades APIs, data model, request flow, the stated load and at least one failure case.
5 of the 22 with a known round come from a design review, where you critique a design instead of drawing one.
7 design questions are reported at both Anthropic and OpenAI. Prep those first if you are in both loops.

The design round runs 55 minutes, and you may write it instead of drawing it

Anthropic’s system design round usually comes on the first onsite day, with coding. A weak first day can cancel the second, so this round decides whether you get to the hiring manager and the values round. It runs 55 minutes with no AI help. You may be asked to type the design into a shared Google Doc instead of drawing it. Some interviewers let you share any drawing tool, but practise writing a design as text.

The round grades five things: your APIs, the data model, the request flow, how you handle the load they state, and at least one failure case. Managers may review a design doc with you instead, and some loops run two design rounds on the team’s own domain. The rest of Anthropic’s loop is in our post on its coding questions and values round.

21 of 29 design questions are about AI systems or ML

We counted every design question reported at Anthropic (n=29): AI system design, ML design and classic system design. AI systems lead with 12. Only 8 are the classic kind you find in a system design book. 5 carry a GPU tag: batched LLM inference API over GPUs, design an LLM inference serving platform, design a distributed AI model downloader, matmul performance modeling on a fixed GPU and design a GPU scheduling platform.

5 of the 9 ML design questions come from a take-home, a practical round or the screen, not from the design round. If you are interviewing for a software engineering role, the AI system questions are the ones to drill.

The 5 reported the most were all asked again since Sep 11, 2026

These 5 are reported at Anthropic more than any of its other design questions, the most reported first:

Batched LLM inference API over GPUs, last asked Oct 15, 2026
Design an AI prompt playground, last asked Oct 13, 2026
One-on-one chat system with presence and offline queue, last asked Oct 13, 2026
Design a distributed AI model downloader, last asked Oct 5, 2026
Critique a design doc from the team's backlog, last asked Sep 11, 2026

3 of the 5 are AI infrastructure. The classic one is a one-on-one chat system with presence and offline queue, a question any company could ask. The fifth asks you to critique a design doc from the team's backlog, which means pulling apart someone else’s design. In all, 9 of Anthropic’s design questions were asked since Sep 1, 2026.

Only 11 of 22 come from the design round itself. 5 come from a design review

Anthropic’s design questions by the round candidates named (n=22 with a known round). The design round produces 11. A design review, where you read a design and critique it, produces 5. Besides critique a design doc from the team's backlog, the reviews cover usage billing for an LLM API, rolling out a new model version, checkpointing a large training run and an eval gate for model launches. The rest come from the screen, practical coding and take-homes.

A review tests the same judgment from the other side. Practise reading a design doc and naming what breaks first, what it costs, and what you would ask the author.

OpenAI’s mix runs the other way. 31 of 50 are classic

OpenAI has 50 reported design questions to Anthropic’s 29, out of 144 questions of any kind to Anthropic’s 72. The share is what differs. 72% of Anthropic’s design questions are AI or ML, against 38% of OpenAI’s. OpenAI’s most reported design questions are design an online chess game, design a payment system and design a video generation service on a scarce GPU pool. Our OpenAI post covers its loop.

7 design questions are reported at both labs: design an AI prompt playground, design an LLM inference serving platform, evaluate a new LLM before launch, design ChatGPT, design a Slack-like chat system, design a RAG system over company docs and design a GPU scheduling platform. If you are in both loops, start there. Our post on AI lab design questions compares nine companies.

What this means for your prep

Start with the 3 AI infrastructure questions reported the most: batched LLM inference API over GPUs, design an AI prompt playground and design a distributed AI model downloader. 21 of 29 design questions are about AI systems or ML, so a classic-only prep leaves most of them uncovered.
Practise in a plain document, 55 minutes on the clock and no AI. Write the API, the data model, the request flow, the load and one failure case.
Keep one classic design ready, starting with a one-on-one chat system with presence and offline queue. Only 8 of Anthropic’s 29 design questions are classic, but one of them is among the 5 reported the most.
Practise reviewing a design. 5 of 22 questions with a known round come from a design review.
In both loops? Do the 7 shared questions first, then OpenAI’s classic set.
Methodology

29 design questions reported at Anthropic and 50 at OpenAI (AI system design, ML design and classic system design), 61 of them with a last-asked date (newest Oct 16, 2026), taken Oct 19, 2026. 22 of Anthropic’s have a known round. “Reported the most” means rated high or very high on how much candidates report them. A question counts as AI or ML when it is an AI system or ML design question; a GPU tag is one of the topic tags we give each question. The round (length, format, AI rule and what it grades) comes from candidate reports and public guides, compiled in Land the Lab’s loop for Anthropic.

Limits: the sample is small. 7 of Anthropic’s design questions have no known round, and 11 have no last-asked date. Reports lean toward software engineering roles. We don’t publish counts per question, only the tier. Loops vary by team and level.

How we made this: Land the Lab counted the reports by type and round. The charts are generated from that data; the text was drafted with AI and checked against it.

Originally published at landthelab.com, where every company's interview loop is mapped round by round.

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

Land the Lab — Mon, 19 Oct 2026 18:53:58 +0000

Photo: Patrick Fore on Unsplash

12 coding questions are reported at OpenAI more than any others, and 10 were asked again since Sep 1, 2026. The coding screen allows no AI and no 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:

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.

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:

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’.

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

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.
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.
For an ML or research role, drill the 4 ML coding questions in plain PyTorch and NumPy.
Practise classic design. 31 of 50 design questions are classic system design. Bring a concrete API and schema, and say how retries stay idempotent.
Write down your view on AI safety before the recruiter call, and pick the one or two projects you will present. Both are gates.
Methodology

144 questions reported at OpenAI, 134 of them with a last-asked date (newest Oct 16, 2026), 136 with a known round, taken Oct 19, 2026. “Reported the most” means the coding questions rated high or very high on how much candidates report them; the next tier is rated medium. 5 coding questions have no rating and are left out of that chart. ML coding means a question tagged machine learning, PyTorch, NumPy, autograd or transformers. The loop (rounds, lengths, formats, AI rules and what each round grades) comes from candidate reports and public guides, compiled in Land the Lab’s loop for OpenAI.

Limits: reports lean toward software engineering roles; ML and research roles get different screens. We don’t publish counts per question, only the tier. Loops vary by team and level, and AI rules change.

How we made this: Land the Lab counted the reports by round and type. The charts are generated from that data; the text was drafted with AI and checked against it.

Originally published at landthelab.com, where every company's interview loop is mapped round by round.

The 5 coding questions Anthropic keeps asking, and the round most people fail

Land the Lab — Mon, 19 Oct 2026 12:00:00 +0000

Photo: Bisworaj Saheb on Unsplash

5 coding questions are reported at Anthropic more than any others, and each was asked again between Sep 18, 2026 and Oct 13, 2026. The round candidates say most people fail has no code.

5 coding questions are reported at Anthropic more than any others, and each was asked again between Sep 18, 2026 and Oct 13, 2026. A crawler, a cache, a file deduper: parts you would build at work.
The coding round produces more reported questions than any other: 19 of 63. Python by default, no AI, and follow-ups on scale and parallelism.
21 of 29 design questions are about AI systems or ML. Inference serving and model distribution, written out in a shared document.
Candidates and recruiters call the values round the one most people fail. It has no code.

Anthropic runs 8 rounds, and a weak first onsite day can cancel the second

The loop starts with a recruiter call and a 90-minute CodeSignal assessment built in levels. Then comes a 55-minute live screen. The onsite spreads over two days: coding and system design usually come first, then the hiring manager, the project deep dive and the values round. The second day only goes ahead after a good first one. Each round and its length:

4 of the 8 rounds put code or a design in front of you, and all 4 rule out AI by default. The screen lets you look up documentation and search the web. Some 2026 loops add a round where you write and review code with Claude Code; there AI use is expected.

The coding round produces more reported questions than any other round

We sorted Anthropic’s 72 reported questions by the round candidates named (n=63 with a known round). Coding leads with 19, ahead of online assessment (9) and ai system design (6).

The same 5 coding questions came back between Sep 18, 2026 and Oct 13, 2026

These 5, from the coding round and the screen, are reported at Anthropic more than any of its other coding questions. Every one was asked again in those weeks:

Fix a function-cache key function, then make it durable, last asked Oct 13, 2026
Concurrent same-host web crawler, last asked Oct 5, 2026
Image transform pipeline with a wall-clock budget, last asked Oct 5, 2026
Deduplicate files by size then content hash, last asked Sep 18, 2026
Reconstruct call stacks from enter and exit traces, last asked Sep 18, 2026

One tier down, 3 more were asked in the same weeks:

Banking system with transfers, cashback and account merges, last asked Oct 13, 2026
Distributed mode and median across worker shards, last asked Sep 18, 2026
Greedy longest-match tokenizer, last asked Sep 18, 2026

Each is a component you would build at work. In the screen, the base problem takes about a third of the hour. The rest is follow-ups: what happens at scale, whether the work is IO-bound or CPU-bound, and how you would run it in parallel. You write Python unless you ask otherwise, and you get no AI help.

21 of 29 design questions are about AI systems or ML

Anthropic’s design questions by type (n=29). The round lasts 55 minutes, and you may be asked to type the design into a shared Google Doc instead of drawing it. It grades your APIs, data model, request flow, how you handle the stated load, and at least one failure case. The prompts lean toward AI infrastructure: batched LLM inference API over GPUs, design an LLM inference serving platform, design a distributed AI model downloader and design an AI prompt playground.

The round most people fail has no code in it

The values round runs 60 minutes, and it can have two interviewers: one asks, one shadows. They want your considered views on AI’s effects on society and on Anthropic’s mission, how you handle ethical conflicts, and whether you change your mind for good reasons. The follow-ups go several levels deep, so a rehearsed story falls apart. Recruiters may send reading on AI safety before the onsite. Read it and form your own view.

The project deep dive sits on the second day too. You present a project you owned for about 20 minutes, and they probe it for the rest of the 60.

Anthropic’s own agents work 6 hours for each staff hour. Its interview still wants your judgment

Anthropic’s report on how employment is changing there (Oct 16, 2026) counts 6 hours of interactive AI agent work for every employee hour by August 2026. Hiring managers’ notes on technical candidates’ AI tool use have tripled since 2024. Read the loop with that in mind. The coding rounds check that you can build without help. The conversations check the judgment you would bring to work where agents do much of the typing.

What this means for your prep

Build the 5 components first, in plain Python and without AI: fix a function-cache key function, then make it durable, concurrent same-host web crawler, image transform pipeline with a wall-clock budget, deduplicate files by size then content hash and reconstruct call stacks from enter and exit traces. Then add the 3 from the next tier.
After each one works, make it concurrent. Practise saying whether it is IO-bound or CPU-bound and what changes at ten times the input.
Practise design in a document. 21 of 29 design questions are about AI systems; write the API, data model and one failure case in prose.
Prepare the values round like a technical one. Read Anthropic’s published writing on AI safety, decide where you agree and where you don’t, and be ready to defend it several follow-ups deep.
Pick your deep-dive project now and cut it to a 20-minute story you can be questioned on for the rest of the hour.
Methodology

72 questions reported at Anthropic, 55 of them with a last-asked date (newest Oct 15, 2026), 63 with a known round, taken Oct 19, 2026. “Reported more than any others” means the coding questions rated high or very high on how much candidates report them; the next tier is rated medium. The loop (rounds, lengths, formats, AI rules and what each round grades) comes from candidate reports and public guides, compiled in Land the Lab’s loop for Anthropic. The figures on agent hours and hiring-manager notes are Anthropic’s own, from its report.

Limits: reports lean toward software engineering roles; research roles choose from a different menu of screens. We don’t publish counts per question, only the tier. Loops vary by team and level, and AI rules change.

How we made this: Land the Lab counted the reports by round and type. The charts are generated from that data; the text was drafted with AI and checked against it.

Originally published at landthelab.com, where every company's interview loop is mapped round by round.

The AI lab interview round that asks the most questions isn't in the onsite.

Land the Lab — Sun, 18 Oct 2026 12:00:00 +0000

Photo: Vitaly Gariev on Unsplash

34% of reported questions at the top AI labs come from the technical screen, more than from any onsite round. 9 loops compared: rounds, question types and AI-tool rules.

34% of reported questions come from the technical screen, more than any onsite round.
4 to 8 rounds; 7 is typical. Cursor, part of SpaceXAI (formerly xAI), runs the shortest loop (4) around a one- or two-day onsite project.
54% of questions are coding. Mistral AI and Cohere are the exceptions: ML design leads there.
AI tools split the labs. Not allowed at OpenAI, Google DeepMind, Perplexity and Cursor; allowed at Cohere.

Seven rounds is typical; Cursor runs 4

Rounds from the first recruiter call to the last interview, as candidates report them. Cursor is part of SpaceXAI (formerly xAI) but hires through its own loop, so it gets its own row.

4 of 9 loops make you build something real

Every loop opens with a recruiter call and all but one have a technical screen. Four have a practical coding round, where you build or fix something real: OpenAI, SpaceXAI (xAI), Scale AI and Cursor (SpaceXAI).

Coding leads everywhere except Mistral AI and Cohere

Question types per company. Google DeepMind is left out: too few reports to chart.

AI tools: not allowed at 4 companies, allowed at 1

What this means for your prep

Treat the screen as the interview. More reported questions are asked there than in any other round; practice multi-part problems that grow every 15 minutes.
Practice building, not just solving. At OpenAI, SpaceXAI (xAI), Scale AI and Cursor (SpaceXAI), a round is about working code you would ship.
Match the mix. Interviewing at Mistral AI or Cohere? Spend more time on ML design than on algorithms.
Ask about AI tools in the recruiter call. Practice both ways if the answer is unclear.
Methodology

328 reported questions across nine loops at eight companies (Cursor is part of SpaceXAI, formerly xAI; one per question and company), 265 of them dated (Jun 7, 2025 to Oct 10, 2026; 151 in the last six months). Loops and AI-tool policies come from candidate reports and public guides, each loop checked in October 2026. Questions without a known round are left out of the round chart.

Limits: reports lean toward software engineering roles, loops vary by team and level, and AI policies change. Google DeepMind had too few reports for the question charts.

How we made this: Land the Lab counted the reports by company, round and type. The charts are generated from that data; the text was drafted with AI and checked against it.

Originally published at landthelab.com, where every company's interview loop is mapped round by round.

OpenAI still asks you to design TinyURL. Mistral asks about mixture of experts.

Land the Lab — Sun, 18 Oct 2026 12:00:00 +0000

Photo: Keawan A on Unsplash

60% of OpenAI's design questions are classic system design, like TinyURL. At Mistral AI, none are. Across 9 AI companies, 60% of design questions are about AI systems and ML.

60% of OpenAI’s design questions are classic system design, like TinyURL, YouTube and a payment system. At Mistral AI, 0 of 12 are.
60% of all design questions are about AI systems and ML. At Anthropic it is 71%.
“Design a RAG system over company docs” comes up at 5 companies, more than any other design question; “Evaluate a new LLM before launch” at 4.
LLM serving and GPUs is the largest AI theme: 13 of 60 AI design questions.
29% of design questions are asked in the technical screen, not only in design rounds.

OpenAI leans classic; Anthropic and Mistral AI lean AI

Each company’s design questions by type. Classic system design is the product kind (TinyURL, a chat app, a payment system); AI system design is serving, RAG and agents; ML design is models, training and evals. Cohere, Perplexity, Scale AI, Cursor (SpaceXAI) and Google DeepMind have fewer than 10 design questions and are left out.

5 companies ask the same RAG question

Design questions reported at more than one company. “Design a RAG system over company docs” comes up at Anthropic, Cohere, Perplexity, SpaceXAI (xAI) and OpenAI.

LLM serving and GPUs lead the AI questions

The 60 AI system and ML design questions, each counted once, by what they are about.

29% of design questions are asked in the technical screen

Design questions by the round they were asked in, where candidates said. The technical screen asks them too.

What this means for your prep

At OpenAI, prep classic system design as well as AI. 60% of its design questions are the classic kind.
Everywhere else, lead with AI systems. The largest themes: LLM serving and GPUs (13), evals and safety (10), training and post-training (9).
Practice the shared questions first: Design a RAG system over company docs, Evaluate a new LLM before launch, Design ChatGPT and Design an LLM inference serving platform.
At Mistral AI, know the model itself. Its questions cover mixture of experts, parallelism, normalization and optimizers.
Have a short design answer ready for the screen. 29% of design questions come up there.
Methodology

126 reports of 105 design questions (classic system design, AI system design and ML design) at 9 AI companies, 94 of them dated (Jul 8, 2025 to Oct 12, 2026). A question asked at two companies counts once for each. Themes come from each question’s tags, one theme per question. Rounds are counted where candidates named them (102 reports).

Limits: reports lean toward software engineering roles, and smaller companies have few design reports. Companies with fewer than 10 are left out of the type chart.

How we made this: Land the Lab counted the reports by company, type, round and theme. The charts are generated from that data; the text was drafted with AI and checked against it.

Originally published at landthelab.com, where every company's interview loop is mapped round by round.

Cursor sells AI coding, but its screens allow only autocomplete. OpenAI's allow none.

Land the Lab — Sun, 18 Oct 2026 12:00:00 +0000

Photo: Joao on Unsplash

OpenAI, Google DeepMind, Cursor and Perplexity keep AI out of every coding round. 8 of 12 AI labs allow it somewhere or are trying it, Anthropic among them. Which rounds, what they grade, and how big tech compares.

Cursor sells AI coding, but its screens allow only autocomplete. OpenAI’s allow none.
4 of 12 AI labs keep AI out of every coding round: OpenAI, Google DeepMind, Cursor and Perplexity.
5 allow it in at least one round (SpaceXAI, Mistral AI, Harvey, Applied Intuition and Cohere); 3 are trying it in some loops: Anthropic, Sierra and Together AI.
Big tech allows AI more: 7 of 8 big tech and other tech companies allow or trial AI, against 8 of 12 AI labs.
Where AI is allowed, those rounds grade explaining what you shipped and a working result more than anything else.

4 of 12 AI labs keep AI out of every coding round

AI labs by their AI policy across the whole loop, where candidates, public guides or the company state one. At Cursor, its CEO says the first technical screens allow no AI other than autocomplete; whether the onsite project allows it is not confirmed. 11 AI labs have no stated policy and are left out.

Big tech allows AI more than the AI labs do

The same policies for big tech and other tech companies (Nvidia, Stripe, Meta, Google, Amazon, Microsoft, Uber and Snowflake). 88% of them allow or trial AI in at least one round, against 67% of the AI labs.

At the AI labs, the coding round is where AI comes in most

The rounds where each AI lab allows AI, or trials it in some loops. A practical round is building or fixing something real.

Those rounds grade explaining what you shipped, then a working result

What the 8 AI lab rounds that allow or trial AI say they grade, by theme. A round can grade more than one thing.

What this means for your prep

Practice unaided first. OpenAI, Google DeepMind, Cursor and Perplexity keep AI out of their coding rounds.
Practice with an assistant too. At SpaceXAI, Mistral AI and Harvey, at least one round allows it.
Ask which rounds allow AI. At 3 labs it depends on the loop, and rules change fast, so ask your recruiter for each round.
Where AI is allowed, get to a working result and be ready to explain every line.
Methodology

The loops of 37 tech companies, round by round, from candidate reports, public guides and company statements, checked in October 2026: 23 AI labs (frontier labs and AI-native companies) and 14 big tech and other tech companies. A company counts as allowing AI when at least one round does; as trying it when only some loops, sessions or pilots do; as keeping AI out when every stated round does (autocomplete aside). Grade themes come from keyword rules over each round’s description.

Limits: policies change fast and differ by team and level, and prep guides contradict each other; where a company has spoken, its word wins. Confirm with your recruiter.

How we made this: Land the Lab sorted each round’s stated policy. The charts are generated from that data; the text was drafted with AI and checked against it.

Originally published at landthelab.com, where every company's interview loop is mapped round by round.

Most AI labs grill you on one past project. Most big tech loops don't.

Land the Lab — Sun, 18 Oct 2026 12:00:00 +0000

Photo: NEOM on Unsplash

12 of 23 AI labs, OpenAI and Anthropic among them, have a round on one project you built, against 3 of 7 big tech companies. What it grades, how long it runs, and how to prepare.

12 of 23 AI labs have a round on one project you built, including Anthropic, OpenAI, SpaceXAI and Google DeepMind.
Most big tech loops don’t: 3 of 7 big tech companies have one, and 2 of 7 scaled startups.
10 of the 12 grade depth under follow-ups: how many layers of “why” your answers survive.
What you personally owned (5) and decisions and tradeoffs (5) come next.
60 minutes is the usual length (6 of the 8 with a known length).
6 ask you to present, often with slides; the rest are conversations.

Most AI labs grill you on one past project; most big tech doesn’t

Loops with a round about a project you built, by kind of company.

10 of 12 AI lab deep dives grade depth under follow-ups

Each deep dive round and what it grades, from its description. The round goes by many names: project presentation, experience deep dive, project walkthrough, take-home review. Microsoft AI doesn't describe what it grades.

The same, by theme. A round can grade more than one thing.

Plan for 60 minutes on one project

Deep dives by length, where it is known. 6 of 12 ask you to present first (Anthropic, OpenAI, SpaceXAI, Google DeepMind, Cursor and Applied Intuition); the rest start as a conversation.

What candidates report about the round

What this means for your prep

Pick one project you owned end to end. 10 of 12 rounds grade depth under follow-ups, so choose the one you know deepest.
Know every decision and the option you turned down. At OpenAI, expect “why not the alternative?” and “what happens at ten times the scale?” for every project.
Say what you did, not what the team did. 5 of 12 rounds grade what you personally owned.
Have a presentation ready for Anthropic, OpenAI and SpaceXAI: about 20 minutes at Anthropic; at OpenAI, bring backup detail beyond the slides.
Keep the story the same in every round. At Decagon, details must stay consistent across interviews.
Methodology

The loops of 37 tech companies, round by round, from candidate reports and public guides, checked in October 2026: 23 AI labs (frontier labs and AI-native companies), 7 big tech companies and 7 scaled startups. 12 of the AI labs include a round about a project you built. Grade themes come from keyword rules over each round’s description; lengths are counted where known (8 of 12).

Limits: the round varies by team and level; at some companies it is reported for many roles, not all.

How we made this: Land the Lab sorted each round’s description. The charts are generated from that data; the text was drafted with AI and checked against it.

Originally published at landthelab.com, where every company's interview loop is mapped round by round.

DEV Community: Land the Lab

72% of Anthropic's design questions are AI or ML. 62% of OpenAI's are classic.

The design round runs 55 minutes, and you may write it instead of drawing it

21 of 29 design questions are about AI systems or ML

The 5 reported the most were all asked again since Sep 11, 2026

Only 11 of 22 come from the design round itself. 5 come from a design review

OpenAI’s mix runs the other way. 31 of 50 are classic

What this means for your prep

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

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

The technical screen produces more reported questions than any other round

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

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

31 of 50 design questions are classic system design

The recruiter call can end your loop over AI safety

What this means for your prep

The 5 coding questions Anthropic keeps asking, and the round most people fail

Anthropic runs 8 rounds, and a weak first onsite day can cancel the second

The coding round produces more reported questions than any other round

The same 5 coding questions came back between Sep 18, 2026 and Oct 13, 2026

21 of 29 design questions are about AI systems or ML

The round most people fail has no code in it

Anthropic’s own agents work 6 hours for each staff hour. Its interview still wants your judgment

What this means for your prep

The AI lab interview round that asks the most questions isn't in the onsite.

Seven rounds is typical; Cursor runs 4

4 of 9 loops make you build something real

More questions come up in the screen than in any onsite round

Coding leads everywhere except Mistral AI and Cohere

AI tools: not allowed at 4 companies, allowed at 1

What this means for your prep

OpenAI still asks you to design TinyURL. Mistral asks about mixture of experts.

OpenAI leans classic; Anthropic and Mistral AI lean AI

5 companies ask the same RAG question

LLM serving and GPUs lead the AI questions

29% of design questions are asked in the technical screen

What this means for your prep

Cursor sells AI coding, but its screens allow only autocomplete. OpenAI's allow none.

4 of 12 AI labs keep AI out of every coding round

Big tech allows AI more than the AI labs do

At the AI labs, the coding round is where AI comes in most

Those rounds grade explaining what you shipped, then a working result

What this means for your prep

Most AI labs grill you on one past project. Most big tech loops don't.

Most AI labs grill you on one past project; most big tech doesn’t

10 of 12 AI lab deep dives grade depth under follow-ups

Plan for 60 minutes on one project

What candidates report about the round

What this means for your prep