In this article
How to use this list
Read the format entries first, because the format explains most of the rest. There is no single format: one guide describes OpenAI's hiring process as a hybrid in which you apply to a specific role and team, and says your mileage may vary depending on team and role (interviewing.io, undated, read October 2026), so your own invitation is the authority on what you will face.
No public frequency data exists and no company publishes a question list, so "most common" here means most often described by candidates, and this page gives no percentages. Treat any list, including this one, as a map of shapes, not a forecast. If you need the method behind a single problem, the coding interview guide teaches the routine, and each entry links a deeper page where one exists. Candidates also report variants, such as a six-part assessment (Hello Interview, undated).
The format AI companies use
What is the Industry Coding Assessment?
The Industry Coding Assessment (ICA) is a CodeSignal test that contains one domain-agnostic, project-based question with four progressive levels (CodeSignal, September 2026). Candidates report that Anthropic uses it, which is a candidate report and not an Anthropic statement (IGotAnOffer, undated). It is one growing project, not four separate puzzles.
Why it is asked: it measures how you extend working code. Read the full explanation: CodeSignal, ICA rules.
How long is the assessment and how is it scored?
CodeSignal's documentation gives a maximum of 90 minutes and a Coding Score range of 200 to 600, where higher scores mean more questions completed successfully (CodeSignal, September 2026). How a score converts to pass or fail is not published, so treat any community "cutoff" as unofficial.
Why it is asked: knowing the scale stops you from guessing at a target. Read the full explanation: CodeSignal, ICA rules.
Do you have to finish all four levels?
No. CodeSignal states that candidates are not necessarily expected to complete all questions within the time limit (CodeSignal, September 2026). The first level asks for basic implementation that accounts for corner cases, so secure the early levels cleanly before reaching for the later ones.
Why it is asked: candidates who rush level one often break it while chasing level four. Read the full explanation: CodeSignal, ICA rules.
Why does backward compatibility matter in the later levels?
Each level builds on the last, so CodeSignal says candidates must reuse, encapsulate and refactor earlier code to keep it backward compatible (CodeSignal, September 2026). A new feature that changes what an old function returns can fail earlier tests. Practice extending code you already wrote while keeping its earlier behavior working (Hello Interview, undated).
Why it is asked: it mirrors real maintenance work. Read the full explanation: CodeSignal, ICA rules.
Knowing it is half the job. Saying it is the other half.
Interview Clozo is an AI interview copilot for real interview rounds: it hears the question and puts a strong answer on your screen, and you say it in your own words. See the features or how it works.
Start Free Trial →Build-it-in-levels questions
What does an in-memory database question look like?
Candidates describe an in-memory database as a representative example (Hello Interview, undated). An illustrative design: level one sets, gets and deletes keys; level two adds a calculation or an export; level three adds expiry times; level four adds a snapshot or rollback. This is a made-up sketch, not the real question. A hash map gives average constant-time get and set.
Why it is asked: it shows whether your first design survives new requirements. Read the full explanation: what defaultdict is in Python, a handy tool for the key-value store at its core.
What are time-based and versioned data structure questions?
One guide lists time-based data structures and versioned data stores among niche topics reported for OpenAI (interviewing.io, undated). The idea is a store where a read depends on a timestamp: a key's value at time t, or the value after an expiry. Keep, per key, an ordered list of (time, value) pairs and search it by time.
Why it is asked: time adds a second dimension to an ordinary map. Read the full explanation: the coding interview guide.
What object-oriented design questions are reported?
For OpenAI, one guide lists abstract classes, iterators and inheritance as niche topics (interviewing.io, undated). Another guide lists candidate-reported coding problems, among them a banking application, described in its own entry below (IGotAnOffer, undated). Expect to model a small domain with clear classes, then extend it without rewriting.
Why it is asked: it separates working code from maintainable code. Read the full explanation: the coding interview guide.
What concurrency questions come up?
One guide lists coroutines and concurrency among niche OpenAI topics (interviewing.io, undated), and a candidate-reported concurrent web crawler is described below (IGotAnOffer, undated). A typical task makes a store or a queue safe when several workers use it. Name the shared state, protect it, and say what you would test.
Why it is asked: shared state is where real systems fail. Read the full explanation: what parallel processing is.
Other problems candidates report
How do you build a web crawler?
A web crawler or concurrent web crawler is a problem reported by candidates on Glassdoor, per IGotAnOffer (IGotAnOffer, undated); one version counts the URLs that match a domain. Keep a queue of pages to visit and a set of visited URLs, fetch a page, keep links on the target domain, and count each once. For the concurrent version, guard the shared set and queue.
Why it is asked: it combines graph traversal with shared state. Read the full explanation: what parallel processing is.
How do you design a banking application?
A banking application with deposits and related operations is reported by candidates on Glassdoor, per IGotAnOffer (IGotAnOffer, undated). Model accounts as objects holding a balance, give each operation one method, reject invalid input such as a negative deposit or an unknown account, and keep earlier methods working as you add transfers or history.
Why it is asked: it tests clean object-oriented design that grows level by level. Read the full explanation: the coding interview guide.
What is the file store question?
A file store with set, get, filter, backup and restore is reported by candidates on Glassdoor, per IGotAnOffer, which lists it under system design (IGotAnOffer, undated). Back it with a hash map from file name to its data and metadata. Filter scans entries by a condition, backup copies the map, and restore swaps the copy back in.
Why it is asked: it is a small key-value store with a snapshot, the same shape as a build-in-levels task. Read the full explanation: what defaultdict is in Python.
Classic data-structure questions
Are classic LeetCode-style problems still asked?
Fundamentals still matter, because the structures under the build questions are hash maps, queues and sorted lists. No public source gives how often AI companies ask classic problems. A widely used baseline for practising them is the LeetCode Top Interview 150 study plan, which is a practice list, not a record of what any company asks.
Why it is asked: it checks your fundamentals under a practical wrapper. Read the full explanation: the coding interview guide.
How do you build a least recently used cache?
A least recently used (LRU) cache is a fixed-size store that evicts the entry unused for the longest time. Pair a hash map with a doubly linked list so that get and put both take constant time: the map finds the node, and the list moves it to the front or drops the tail. It is a staple of practice lists such as the Top 150 plan above.
Why it is asked: it combines two structures behind one interface. Read the full explanation: the coding interview guide.
How do you solve longest substring without repeating characters?
Slide a window over the string, keep the characters inside it in a set or map, and move the left edge past the earlier copy when a repeat appears. Each character enters and leaves the window once, so the run is linear.
Why it is asked: it is the standard sliding-window test. Read the full explanation: longest substring without repeating characters.
Is dependency ordering asked, and how do you solve it?
Ordering tasks by dependency is a classic graph task, and the method is a topological sort: repeatedly take a node with no unmet dependencies, and if none remain while nodes are left, there is a cycle. Treat it as practice for any build-order or scheduling problem.
Why it is asked: it tests graph modeling on a realistic task. Read the full explanation: topological sort explained and detecting a cycle in a directed graph.
Which patterns sit under these questions?
Three recur in practice, whichever company asks. A hash map with running totals solves counting questions, see subarray sum equals k. A heap or a sort handles "top few" questions, see k closest points to origin. Sorting then sweeping handles ranges, see merge intervals. This is practice advice, not a claim about any one company.
Why it is asked: patterns let you recognize a new problem quickly. Read the full explanation: the coding interview guide.
Language, tools and rules
Which programming language should you use?
Python is common in candidate reports, though language requirements vary by role (Hello Interview, undated). Whether a company limits the choice is not published, so check your invitation. Choose the language in which you can write correct, tidy code fastest.
Why it is asked: it is a setup choice, not a trick. Read the full explanation: Python interview questions and answers.
Can you use AI during the interview?
Check the invitation, because rules differ by step. Anthropic's guidance, as reproduced by a newsletter, says live interviews are "no AI assistance unless we indicate otherwise" (craai newsletter, read October 2026). Anthropic's own careers page is the authority; confirm the rule there. For OpenAI, no first-party rule was found in the sources read, so follow your invitation.
Why it is asked: breaking a stated rule can end your candidacy. Read the full explanation: Anthropic guidance, as reproduced.
Does a top score mean you advance?
Not necessarily. A guide reports that a high score is necessary but not sufficient, and candidates report near-perfect scores that did not advance (IGotAnOffer, undated). That is secondary reporting, and the companies publish no pass rate. Prepare for the later interview rounds as well as the assessment.
Why it is asked: it corrects a common belief about scores. Read the full explanation: the coding interview guide.
A worked example: how four levels grow
CodeSignal's technical brief describes the first level as basic programming with simple data structures, taking about 10 to 15 minutes, and gives 30 to 60 minutes for each of levels 3 and 4 (CodeSignal brief, undated). Only the level 2 budget, 20 to 30 minutes, is an illustrative figure. The four budgets add up to 90 to 165 minutes.1 The low estimates alone add up to 90 minutes, which fills the whole cap. The high estimates add up to 165 minutes, which exceeds the cap by 75 minutes.
| Level | Budget in minutes | Source status |
|---|---|---|
| 1 | 10 to 15 | CodeSignal brief |
| 2 | 20 to 30 | illustrative |
| 3 | 30 to 60 | CodeSignal brief |
| 4 | 30 to 60 | CodeSignal brief |
| Total | 90 to 165 | computed |
Text description of the diagram
- The low estimates alone add up to 90 minutes, which fills the whole cap.
- The high estimates add up to 165 minutes, which exceeds the cap by 75 minutes.
The same script runs two toy traces, both illustrative. A store with an expiry time: set key b to 2 at time 0 with a lifespan of 5, read it at time 3 and get 2, read it at time 6 and get nothing. A snapshot is taken at time 6, then a is overwritten from 1 to 9, and restoring the snapshot brings a back to 1. A least recently used cache of size 2: put a, put b, read a, put c, and the entry evicted is b, because a was used more recently.1
Follow-up questions
What are the most common coding interview questions at AI companies?
Candidates describe a few shapes: a multi-level build such as an in-memory database, object-oriented problems such as a banking application, concurrent problems such as a web crawler, and classic data-structure problems. All of it is candidate-reported, so check your own invitation for the exact format.
How should you prepare for an AI company coding round?
Practice extending code you already wrote without breaking earlier behavior, rehearse one hash-map store and one cache from scratch, and revise concurrency basics. Work through the coding interview guide and keep the language you know best.
Does OpenAI ask LeetCode questions?
No public source gives that answer. One guide lists coroutines, concurrency and object-oriented topics as niche OpenAI topics (interviewing.io, undated), and the classic practice list is the LeetCode Top Interview 150 study plan. Prepare fundamentals and read your invitation.
Is there an AI-allowed coding round?
Sometimes the invitation says so. Anthropic's guidance, as reproduced by the newsletter, says "We'll be clear when AI is allowed (example: 'You may use Claude for this coding challenge')" (craai newsletter, read October 2026). Without such a statement, assume no AI assistance.
- AI companies publish no question lists; everything here is candidate-reported.
- CodeSignal's assessment is one project in four cumulative levels, 90 minutes, scored 200 to 600.
- You are not expected to finish every level; keep earlier levels correct and backward compatible.
- Reported problems include in-memory stores built in levels, a banking application, a concurrent web crawler and a file store.
- AI use rules differ by step, so read your invitation before you rely on any tool.
Sources
- What are the Industry Coding Assessment (ICA) rules? (CodeSignal Knowledge Base) — CodeSignal (vendor documentation, accessed 2026-10-26), 14 Sep 2026
- Industry Coding Skills Evaluation Framework Technical Brief — CodeSignal (vendor brief, undated, accessed 2026-10-26), 26 Oct 2026
- OpenAI's Interview Process and Questions — interviewing.io (candidate-report guide, undated, accessed 2026-10-26), 26 Oct 2026
- Anthropic Software Engineer Interview — IGotAnOffer (candidate-report guide, undated, accessed 2026-10-26), 26 Oct 2026
- Anthropic Software Engineer Interview Guide — Hello Interview (candidate-report guide, undated, accessed 2026-10-26), 26 Oct 2026
- Anthropic guidance on candidates' AI usage (as reproduced) — craai newsletter (reproduction of the guidance, accessed 2026-10-26), 26 Oct 2026
- LeetCode Top Interview 150 study plan — LeetCode (study plan, undated, accessed 2026-10-26), 26 Oct 2026
