We tracked 40 questions OpenAI asked in October. Only 5 were about AI.

Photo: Randy Fath on Unsplash Only 5 of the 40 questions OpenAI was reported asking in October were about AI. The rest were mostly coding and classic design: a chess game, a payment system, poker hands. Anthropic: 3 of 12. 5 of the 40 questions OpenAI was reported asking in October were about AI (13%). Over every date we hold it is 26 of 137 (19%), so October was no exception. OpenAI had 40 of the 123 October questions at the 46 AI companies we track, more than any other. Anthropic came next with 12. The technical screen produced 15 of OpenAI’s 40, more than any other round. The online test’s 6 were puzzles, none about AI. 27 of OpenAI’s 40 are reported at no other company we track. A shared question bank won’t cover them. 3 of Anthropic’s 12 were about AI, all of them design questions. Its coding was mostly practical: a bank, a crawler, a cache. OpenAI had 40 of the 123 October questions at the AI companies we track We counted every question at the 46 AI labs and AI companies we track whose last reported date falls in October 2026. Our reports for the month run from Oct 1, 2026 to Oct 29, 2026. OpenAI had 40 of them, more than the next four companies together (37). Anthropic had 12. Companies with fewer than 2 are left out of the chart. A month of reports is a small window, and a quiet month can mean nobody wrote one up. SpaceXAI (formerly xAI) had 13 in September and none in October. Mistral AI had 1 in September and none in October. Treat a company’s October as a sample of what it asks now, not its whole bank. Only 5 of OpenAI’s 40 were about AI We call a question about AI when it is AI system or ML design, or when its topics include machine learning, LLMs, transformers or GPUs. By that count, OpenAI’s October had 5: find noisy annotators and retrain a classifier without them, design an AI chatbot app, design an LLM inference serving platform, causal hard attention with a KV cache and find the planted bugs in a small Transformer training script. The other 35 were 24 coding problems with no AI in them, 9 classic system design questions, one object-oriented design and one behavioral round. This is not a new turn. Over every date we hold, 26 of OpenAI’s 137 questions are about AI (19%). The company that makes ChatGPT mostly tests whether you can write and design ordinary software. Anthropic’s October leaned further toward AI, 3 of 12 (25%). The technical screen produced 15 of OpenAI’s 40, more than any other round OpenAI’s October questions by the round candidates named. The technical screen is the first live round, and it allows no AI and no autocomplete. The screen’s problems were framed as small programs: find noisy annotators and retrain a classifier without them, three-card poker: classify hands and pick the winner, deal five-card poker hands and pick the winners, rectangle board with remove-at-cell, batch remove and undo, follow one-way links to the endpoint or the last node before a loop, count connected friend groups, memory allocator with malloc, free and coalescing, paint house: cheapest daily colors with no repeat on consecutive days, threshold infection spread on a grid, validate chess moves from the starting position, linked list with checkpoints for fast indexed insert and get, causal hard attention with a KV cache and three-card showdown between two players. Some are textbook graph and dynamic programming problems with a story on top, like the friend groups and the house painting. Others are small programs you build in parts: the memory allocator, the poker hands, the chess move checker. The 6 online test questions were the closest thing to classic puzzles: longest increasing subsequence of a price series, price a macaron order with set discounts, shortest path on a grid with a custom move set, simplify a quadratic expression, count ordered routes that visit one shop of each of four types and microorganisms eating neighbours: the largest survivor. Candidates report the timed test mainly for early-career roles. OpenAI’s 9 classic design questions included a chess game, a payment system and a cloud IDE No URL shortener this month. The classic design questions asked for whole products and platforms, newest first: Design a sandboxed cloud IDE, last asked Oct 16, 2026 Design a real-time multiplayer typing race, last asked Oct 14, 2026, in the technical screen Design an online chess game, last asked Oct 14, 2026 Design a payment system, last asked Oct 13, 2026 Design a distributed power grid control system, last asked Oct 12, 2026 Design a real-time trivia battle royale with live spectators, last asked Oct 10, 2026 Design a CI/CD system, last asked Oct 7, 2026 Design an image sharing service that stores each file once, last asked Oct 7, 2026 Design a metrics monitoring system, last asked Oct 2, 2026, in the technical screen The games share a problem. A chess game, a typing race and a trivia battle royale all need live state pushed to several players at once, and a plan for a player who drops. If you can design one of them well, the other two are variations. 27 of OpenAI’s 40 are reported at no other company we track We checked each October question against the other companies in our data. 27 of OpenAI’s 40 (68%) show up nowhere else, against 6 of Anthropic’s 12. The ones OpenAI shares include the familiar designs, like the payment system and the metrics monitoring system. Among the ones only OpenAI asks are the poker hands, the persistent key-value store and the work queue with leases. That changes how you practise. A big list of common problems teaches the patterns, but you won’t meet these exact prompts there. Work from OpenAI’s own list, newest first. Anthropic’s 12: 3 AI design questions, 5 coding problems Anthropic’s October was smaller and split cleanly. The AI questions were batched LLM inference API over GPUs, an AI prompt playground and a distributed AI model downloader, all three of them design questions. Its one classic design question was the one-on-one chat system with presence and offline queue. The five coding problems were concurrent same-host web crawler, banking system with transfers, cashback and account merges, fix a function-cache key function, then make it durable, randomness from a biased coin: fair flips, uniform 1 to n, Monte Carlo and image transform pipeline with a wall-clock budget. The bank, the crawler, the cache and the image pipeline are small pieces of a real system, each harder as the interviewer adds parts. The biased coin is the one puzzle. The other three were conversations: recruiter screen on why Anthropic, how do you prioritize when you have too much work? and project deep dive: present a project you led. What this means for your prep For OpenAI, prep software first and AI second. 35 of its 40 October questions had no AI in them. Get fast at building a small program in parts, with tests, without autocomplete. Practise one real-time game design end to end. The chess game, the typing race and the trivia battle royale share live state, fan-out and reconnects. Still prepare the AI topics. October’s five AI questions were find noisy annotators and retrain a classifier without them, design an AI chatbot app, design an LLM inference serving platform, causal hard attention with a KV cache and find the planted bugs in a small Transformer training script. ML roles get more of these. For Anthropic, split your time. AI system design for the design round, practical coding for the coding rounds, and a story about a project you led. Work from each lab’s own list. 27 of OpenAI’s 40 are asked nowhere else we track. See OpenAI and Anthropic, newest first. Methodology Every question at the 46 AI labs and AI companies we track whose newest reported date at that company falls in October 2026 (123 in all; the newest is Oct 29, 2026), taken Nov 1, 2026. A question asked at two companies counts once at each. OpenAI n=40, Anthropic n=12. A question is about AI when it is AI system or ML design, or tagged with one of these topics: machine learning, PyTorch, NumPy, autograd, transformers, LLMs, LLM serving, KV cache, attention, GPUs. Rounds are the ones candidates named. “Reported at no other company” means no other company in our data has the question. Limits: the month’s reports stop at Oct 29, 2026, so the last days of October are missing. A month is a small sample, and a company with no October questions may simply have had no reports. A question counts in October only when October is its newest date, so questions asked in October and again later would move out. Reports lean toward software engineering roles. How we made this: Land the Lab counted candidate reports by company, type, round and date. 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.