Anthropic Interview Internal Question Bank Exposed: 5 Coding Problems Asked Repeatedly, and the Round You’re Most Likely to Fail

I. Process Overview: 8 Rounds, Two-Day Onsite — Day Two Depends on Day One

1. Recruiter call
2. CodeSignal OA: 90 minutes, progressive leveled problems
3. Live screen: 55 minutes; docs and web search allowed
4. Onsite Day 1: coding + system design
5. Onsite Day 2: hiring manager, project deep dive, values round — if you’re weak on Day 1, Day 2 may be cancelled outright

Key rules:

• Of the 8 rounds, 4 require writing code or doing design. AI is disabled by default in all of them; phone screens allow doc lookup and web search.
• The 2026 loop adds a new round: “Write code with Claude Code and do a code review” — that round actually requires using AI.

II. Top 5 Frequent Problems

Common trait: All are components you’d actually build at work — not LeetCode-style algorithm questions. In phone screens, basic problems take only about 1/3 of the time; the rest is all follow-ups: what if the scale grows, is it IO-bound or CPU-bound, how do you parallelize. Hand-written Python by default, no AI.

1. Fix a function-cache key function, then make it durable

What it tests:

• Cache key correctness (argument normalization, hash collisions, non-hashable arguments)
• Persistence (serialization, writing to disk, recovery, concurrent writes)

2. Concurrent same-host web crawler

What it tests:

• BFS + thread pool/coroutines, URL deduplication, polite rate limiting
• Cross-referenced with the existing “LC webcrawler variant” in the question bank

3. Image transform pipeline with a wall-clock budget

What it tests:

• Batch-processing images × transform combinations per a JSON recipe
• The question bank already has a public outline of this problem (3 parts / 55 min); the deep-dive direction is scalability redesign under large data volumes and expensive transforms

4. Deduplicate files by size, then content hash

What it tests:

• Two-level filtering to cut cost: size grouping → header sampling hash → full hash
• Handling large files and hash collisions

5. Reconstruct call stacks from enter and exit traces

What it tests:

• Diffing profiler sample stacks to produce start/end events — finding the common prefix, closing old frames before opening new ones, careful about nesting order

III. Next Most Frequent 3 Problems

1. Banking system with transfers, cashback, and account merges

What it tests: OOP + state consistency: transfer atomicity, idempotency, balance and cashback rules when merging accounts

2. Distributed mode and median across worker shards

What it tests:

• Distributed aggregation: mode via count merging; median requires histograms/bucketing or approximate algorithms
• The point: you can’t just pull all the data back locally

3. Greedy longest-match tokenizer

What it tests:

• Trie/prefix matching, vocabulary construction, edge cases of longest-match

IV. Design Round: 21 of 29 Problems Are AI Systems / ML

• Duration: 55 minutes; you may be required to type your design into a shared Google Doc rather than draw diagrams.
• Evaluation points: API design, data model, request flow, how to handle the given load, and at least one failure case.
• Typical problems: batched LLM inference API on GPUs, LLM inference serving platform, distributed AI model downloader, AI prompt playground.

V. The Round You’re Most Likely to Fail Has No Code: The Values Round (60 Minutes)

• May have two interviewers: one asks questions, one shadows and observes.
• Tests your genuine views on AI’s societal impact and Anthropic’s mission, how you handle ethical conflicts, and whether you’ll change your position for good reasons.
• Follow-ups dig several layers deep — rehearsed stories get exposed quickly. The recruiter may send AI safety reading materials before the onsite — read them and form your own opinions, including where you disagree.

Project deep dive (Day 2, 60 minutes): Spend about 20 minutes presenting a project you led yourself; the rest is a deep dive. Choose the project now and distill it into a story that can survive 40 minutes of follow-up questions.

VI. A Background Signal: Why the Interview Is Designed This Way

By August 2026, inside the company, every 1 hour of employee work time corresponds to 6 hours of interactive AI agent time. Hiring managers’ records of technical candidates’ AI tool usage have grown 3× since 2024. The original article’s interpretation: the coding round tests whether you can build without assistance; the conversational rounds test the judgment you bring to a role where an agent types for you.

VII. Prep Checklist (Matching the Report’s 5 Recommendations)

• Practice the 5 components first: written in pure Python, no AI. Then practice the next 3 frequent problems.
• Practice a concurrent version of each component: be able to explain IO-bound vs. CPU-bound, and what changes if input volume scales 10×.
• Practice design problems in a document: API, data model, one failure case — all written out in prose. Focus on AI infrastructure topics.
• Prepare the values round like a technical round: read Anthropic’s public AI safety articles, decide what you agree and disagree with, and be able to withstand multiple layers of follow-up.
• Choose your deep dive project in advance: the standard is 20 minutes to present, 40 minutes of deep questioning.