Over half of Google's coding questions are puzzles. At Anthropic, 2 of 29.
We sorted 1,887 interview questions candidates reported at 57 companies in 2026. 53% of Google's coding questions were algorithm puzzles. At Anthropic, 2 of 29.

- 91 of Google’s 172 coding questions in 2026 were classic algorithm puzzles (53%). At Anthropic it was 2 of 29. Across the frontier labs, 11%; across big tech, 45%.
- The frontier labs’ most common coding topic is simulation. They ask you to build a small working system in parts: a bank ledger, a crawler, a work queue.
- 23% of the frontier labs’ 2026 questions touch LLMs, against 7% at big tech. Models, training and GPUs come up most, then agents and tool use.
- 42% of the frontier labs’ and AI companies’ questions were last asked in the past 90 days, against 23% at big tech. A question bank from last year misses much of it.
- First call to offer runs from 1 to 2 weeks at Mercor to 4 to 8 at OpenAI. The table below has each AI company.
Over half of Google’s coding questions are puzzles. At Anthropic, 2 of 29.
We took every question candidates reported in 2026 at 57 companies: 5 frontier labs, 38 AI companies and 14 big tech companies. That is 1,887 questions, last asked between Jan 1, 2026 and Nov 11, 2026. Then we sorted the coding questions into two piles. A classic algorithm puzzle is one whose topics are all algorithms and data structures: arrays, strings, trees, graphs, dynamic programming. Everything else asks you to build a piece of a real system, like a bank, a scheduler or a parser.
At Google, 91 of 172 coding questions were puzzles. At Anthropic, 2 of 29. At OpenAI, 10 of 73. The chart shows every company with at least 15 coding questions this year; the frontier labs keep the colour.
By group: the frontier labs 11% (14 of 133), the AI companies 21% (50 of 238), big tech 45% (379 of 846). Over half at Amazon, Google and Uber. Stripe sits with the labs, at 12%. Its loop is built around practical rounds too.
The labs ask you to build things: simulation leads their coding topics
Each question carries topic tags. Here are the most common on 2026 coding questions, as a share of each group’s coding questions (frontier labs n=133, big tech n=846). A question can have several. The language tag and the general “design” tag are left out.
| Topic | Frontier labs | Big tech |
|---|---|---|
| simulation | 15% (20) | 7% (58) |
| graphs | 9% (12) | 10% (83) |
| hash maps | 8% (11) | 9% (73) |
| concurrency | 8% (11) | 3% (28) |
| object-oriented design | 8% (11) | 2% (20) |
| debugging | 8% (10) | 2% (15) |
| arrays | 5% (6) | 12% (102) |
| BFS | 5% (6) | 10% (87) |
| strings | 3% (4) | 13% (109) |
| trees | 2% (2) | 10% (86) |
Simulation and concurrency mean a program with state that changes over time, built up in parts as the interviewer adds rules. The newest of the frontier labs’ 119 non-puzzle coding questions this year: rectangle board with remove-at-cell, batch remove and undo (OpenAI), concurrent same-host web crawler (Anthropic), three-card poker: classify hands and pick the winner (OpenAI), find noisy annotators and retrain a classifier without them (OpenAI), per-user chat event counts in a sliding window (OpenAI), deal five-card poker hands and pick the winners (OpenAI), fault-tolerant work queue with leases and retries (OpenAI) and speculative edits: decode a code edit using the original file as the draft (SpaceXAI).
The onsite coding round produced 54 of them, more than any other round, and the technical screen 34. The OpenAI technical screen and the Anthropic technical screen pages list what each one asks.
23% of frontier lab questions touch LLMs. At big tech, 7%.
A question touches LLMs when its topics include LLMs, inference, retrieval, agents, evals, training, transformers or GPUs. That covers AI system design and ML design, and coding questions like a KV cache or a batching loop. 51 of the frontier labs’ 219 questions qualify, 82 of the AI companies’ 397, and 93 of big tech’s 1,271.
So 77% of frontier lab questions are not about LLMs at all. The labs still test ordinary software first. When an LLM question does come, it comes early: the technical screen produced 15 of the frontier labs’ 51, more than any other round.
Models, training and GPUs lead the LLM questions, then agents and tool use
The 133 LLM questions at the frontier labs and AI companies, by topic. A question can count under more than one.
- Models, training and GPUs (53): MNIST classifier: per-example loss bounds and adaptive calibration error (OpenAI), locate a point light source from photon hits on a line by MLE (OpenAI) and find noisy annotators and retrain a classifier without them (OpenAI).
- Agents and tool use (25): design an evaluation platform for a design-editing AI agent (OpenAI), multi-step task with an LLM tool-use loop (Anthropic) and fix an LLM agent, then cut its turns to one (Anthropic).
- Inference and serving (20): design an LLM inference API that batches requests on GPUs (Anthropic), speculative edits: decode a code edit using the original file as the draft (SpaceXAI) and design an LLM inference serving platform (OpenAI).
- RAG and retrieval (20): ML design for retrieval: embeddings, hybrid search and ranking (OpenAI), build a RAG agent on the Mistral API, then defend its design (Mistral AI) and mine new information from a huge unlabeled image corpus (OpenAI).
- Evals (13): A model about to ship regressed on safety evals (Google DeepMind), design an evaluation platform for a design-editing AI agent (OpenAI) and design an agent harness and its evaluation system (OpenAI).
Agents and evals show up at application companies like Harvey, Cohere and Databricks, not only at the labs. Inference questions go deep: a KV cache, continuous batching, speculative decoding.
42% of the labs’ and AI companies’ questions were last asked in the past 90 days
Of every question we hold for the frontier labs and AI companies, of any date, 353 of 839 were last reported asked on or after Aug 15, 2026: 46% at the frontier labs and 40% at the AI companies. At big tech it is 23% (464 of 2,004). Big tech’s lists are larger and reach further back. The labs’ lists are newer, so a question bank from last year will miss much of what they ask now.
First call to offer: 1 to 2 weeks at Mercor, 4 to 8 at OpenAI
Every frontier lab and AI company with at least 10 questions in 2026. Weeks are the range in each company’s loop summary, from candidate reports and public guides.
| Company | Questions in 2026 | Puzzles among coding | Touch LLMs | Weeks to offer |
|---|---|---|---|---|
| OpenAI | 123 | 10 of 73 | 23 | 4 to 8 |
| Nvidia | 52 | 5 of 34 | 7 | 3 to 8 |
| Anthropic | 51 | 2 of 29 | 14 | 3 to 8 |
| Databricks | 46 | 6 of 24 | 4 | 2 to 6 |
| SpaceXAI (formerly xAI) | 36 | 1 of 27 | 6 | 2 to 3 |
| Mercor | 28 | 3 of 11 | 6 | 1 to 2 |
| Applied Intuition | 27 | 5 of 22 | 2 | 2 to 4 |
| Nuro | 27 | 8 of 18 | 3 | not reported |
| Glean | 24 | 7 of 20 | 2 | 2 to 5 |
| Harvey | 20 | 0 of 10 | 7 | 2 to 5 |
| Decagon | 15 | 0 of 7 | 5 | 3 to 4 |
| Scale AI | 15 | 1 of 8 | 5 | 2 to 6 |
| CoreWeave | 13 | 0 of 5 | 0 | 3 to 5 |
| Cohere | 10 | 1 of 4 | 6 | 4 to 7 |
| Perplexity | 10 | 1 of 8 | 2 | 3 to 5 |
123 of OpenAI’s questions were last asked in 2026, more than at any other AI company. Each company page maps its loop round by round, with every question newest first.
What this means for your prep
- Stop grinding puzzles for the labs. Only 11% of the frontier labs’ coding questions were classic puzzles in 2026. Practise building a small system in parts instead: a bank with transfers, an in-memory database with TTLs, a crawler with threads. Write tests as you go.
- Keep puzzles for big tech. 53% of Google’s coding questions were puzzles. If you are interviewing at both, split your coding time.
- Learn concurrency in your interview language. Threads, locks and queues are in 8% of the frontier labs’ coding questions, against 3% at big tech.
- Prepare one LLM design end to end. Pick from the topics above: they cover 108 of the 133 LLM questions; the rest are general LLM questions. Know how a KV cache and batching change latency and cost.
- Work from this year’s list. 42% of the labs’ questions are from the last 90 days. Start with OpenAI and Anthropic, newest first.
See every reported question for your loop, by round, with when it was last asked.
Pick your companyWhat the labs asked this week
A weekly email of what OpenAI, Anthropic and the other top AI labs asked engineers, round by round.