Anthropic Interview Questions
Reconstructed from 402 verified candidate reports across 44 questions. Nov 2024 – May 2026.
This page is a live view of every Anthropicinterview question AceOffer has indexed — pulled from real candidate reports, not invented from job descriptions or one founder’s memory. Every question shows how many times it’s been reported and when it was last seen. The catalog gets a refresh pass every month.
The Anthropic loop, from candidate reports
Anthropic's loop is recruiter screen → CodeSignal-style 90-minute Python coding assessment → 1–2 phone screens → a 5-round virtual onsite. The behavioral round at Anthropic is unusually deep and weighted: the interviewer probes AI safety alignment with specific hypothetical scenarios, and candidates report being asked 'why Anthropic over OpenAI' with the expectation of a substantive answer that goes beyond mission slogans. The Web Crawler phone screen has shown up in 57 reports — it's effectively a filter. For ML / Research Engineer roles, the MLE round combines an inference-API system design with deep questions on tokenizer implementation, image-processing pipelines, and Python-from-scratch implementations.
What gets asked, by round
Counts reflect distinct questions per round, not number of times asked. Frequencies on individual question cards show how many candidates reported getting that specific question.
60 minute design rounds. Interviewers push hard on the specific dimension their team cares about (storage at scale, real-time fan-out, multi-tenancy).
Two-way conversations. Anthropic in particular probes AI safety alignment hard; OpenAI probes mission-fit and shipping velocity.
Implement forward + backward from scratch (NumPy), debug a planted-bug transformer, or design an ML system. Math + code + system thinking.
Take-home or proctored 90-minute online assessment before the loop. Used as a filter — not weighted in the final decision once you're in the onsite.
60–75 minute live coding rounds. Multiple sub-problems progressing in difficulty. Test harness usually provided.
60–90 minute coding or system design over CoderPad or similar. Pass bar is correctness + clean communication; brilliance isn't required.
Most reported Anthropic questions
Sorted by candidate-report frequency. These are the questions that have recurred most across the loops we’ve indexed.
| Question | Round | Reported | Last seen |
|---|---|---|---|
| Web Crawler | Phone Screen | 57× | April 2026 |
| File Deduplication | Coding | 44× | April 2026 |
| Profiling / Stack Trace Analysis | Coding | 34× | April 2026 |
| Inference API System Design | System Design | 30× | April 2026 |
| Why Anthropic? | Behavioral | 17× | April 2026 |
| Most Impactful Project / HM Behavioral | Behavioral | 16× | April 2026 |
| AI Safety Culture Fit | Behavioral | 16× | March 2026 |
| Culture Fit / Behavioral Questions | Behavioral | 14× | March 2026 |
| In-Memory Database | OA | 14× | January 2026 |
| Project Deep Dive / Technical Presentation | Behavioral | 14× | April 2026 |
Want to see all 44? Browse the full Anthropic catalog →
Read two Anthropic questions free
Full problem statements, candidate-reported follow-ups, and walkthroughs. No signup needed.
Crawl every page within the same hostname using a provided link helper. Start single-threaded, then make it concurrent. The most-reported Anthropic question by a wide margin.
Design a prompt playground — stateless conversation turns, share-conversation feature, very large prompts. The interviewer probes UX, frontend state, and the share mechanism in depth.
Live-debug a buggy GRPO training script. Three planted bugs — softmax before multinomial, epsilon-less std, and the importance-sampling ratio formula — then RL-theory follow-ups on PPO clipping and mini-epoch policy drift. The Anthropic Research / MLE signature debug round.
- •On the behavioral round: come with one specific Anthropic alignment paper or recent decision (RSP, Constitutional AI, the deprecation policy) and connect it back to your own work — generic safety enthusiasm gets flagged
- •On Web Crawler: progressive complexity wins — solve single-threaded clearly first, then propose the concurrent extension with the interviewer's blessing. Don't dive into concurrency immediately
- •On Inference API system design: distinguish the routing layer (cheap, stateless) from the model serving layer (expensive, GPU-pinned) early, then build from there
- •For the MLE round: have NumPy implementations of softmax + cross-entropy and backprop through MLP cold — the interviewer expects you to derive without looking things up
- •Quote your work back as 'I made this technical decision because of X tradeoff' — not 'I worked on Y' — Anthropic's deep-dive round is grading reasoning, not titles
- •Treating the AI safety culture round as a soft round and giving a corporate-policy answer — gets flagged as 'low alignment signal' and reported as a primary loss reason
- •On Web Crawler: jumping to a concurrent solution without verifying the single-threaded version against the test cases first
- •Skipping the URL normalization step on Web Crawler (fragments, query params, trailing slashes) — multiple candidates lost the round on this single edge case
- •On Inference API: conflating the model-routing problem with the model-serving problem; not distinguishing latency-sensitive (real-time) from throughput-optimized (batch) workloads
- •On the project deep-dive: presenting WHAT you built without articulating the alternatives you considered and why you rejected them
Get the full Anthropic catalog
Every question. Every candidate-reported follow-up. The AI mock interviewer to drill them with. Monthly refresh.