Anthropic Software Engineer Interview Experience — A Community-Compiled Guide to the OA, Technical, and Values Rounds

Anthropic's SWE interview process breaks into three main parts: an online assessment (OA), one to two technical interviews, and a values/culture interview. This guide compiles experiences and community info from multiple candidates, in the hope that it helps people who are preparing.

Timeline

The whole process moves pretty slowly. Multiple candidates report gaps of 2-3 weeks between rounds. A typical timeline: 2-3 weeks after applying for a response, then the OA gets sent, then a recruiter call gets scheduled 3-7 days after you submit it, and after that the technical interviews get scheduled 2-4 weeks later. A long gap doesn't mean something's wrong with your performance — keep preparing, and don't pause your other applications.

Online Assessment (timed, Python)

The OA is a multi-stage coding challenge, usually with 4 levels of increasing difficulty. You can see the full problem from the start, and you work through it in order. Python is the standard language. This isn't leetcode — it's actual system building. Each stage builds on the previous one, so write clean, modular code from the start, or you'll be refactoring under time pressure by stage 3-4.

Reported questions:

• In-memory database (4 levels): Level 1: basic get, set, delete. Level 2: search and filter. Level 3: TTL auto-expiration. Level 4: time-travel queries to look up historical data.
• Cloud storage system (4 stages): Stage 1: add, copy, read files. Stage 2: pattern-matching file search. Stage 3: user management and storage limits. Stage 4: compress and restore files.
• Banking system (4 levels): Level 1: create account, deposit, transfer. Level 2: rank by spending. Level 3: scheduled payments and cashback. Level 4: merge accounts and look up historical balances.
• Employee management system (4 parts): Part 1: add employees, track arrival/departure. Part 2: calculate hours worked, find top performers. Part 3: handle promotions that take effect in the future. Part 4: add double-pay periods.
• LRU cache bug fix: Given a Python LRU cache implementation, there's a bug in how the key is built from *args and **kwargs. Find the bug, fix it, and explain what was wrong.
• Batch image processor: Batch-edit images in a folder according to a JSON config.
• Infection simulation: Simulate infection spread on a 2D grid, where a cell gets infected once K of its neighbors are infected. Needs to be optimized to better than O(NM) per step, using BFS.

You don't need a perfect score to pass. They care more about how far you got and your code quality. Read through each stage's full requirements before you start coding.

Technical Interview

The technical interviews focus on systems and infrastructure. Usually there's no single correct answer. How you talk through your thinking matters as much as your final solution.

Reported coding questions:

• Tokenizer implementation: Implement a tokenizer using a longest-match algorithm. Part 1: basic vocabulary lookup. Part 2: handle unknown tokens. Part 3: extended features like merging consecutive unknown tokens.
• Tokenize/detokenize bug fix: Given buggy tokenize and detokenize functions, read and understand them, then find and fix the bug.
• Token usage calculation: Calculate token costs for an LLM API. Handle different prices for input vs. output tokens, and token counting for streaming responses.
• Call stack sampling to trace events: Convert call stack snapshots from a sampling profiler into an event timeline.
• Distributed mode lookup: Find the most frequent value across data distributed over 10 workers. Constraint: network bandwidth is the bottleneck, so you can't send all the data to one machine.

Reported system design questions:

• Distributed model deployment: Design a system to download and deploy a 500GB ML model to 100-1000 GPU workers. 10 Gbps external bandwidth, 10 Gbps per worker. Minimize deployment time. Handle worker failures. Discuss pipeline vs. tree distribution, chunking, and scaling to 10k workers.
• Inference API (high concurrency): Design a high-concurrency LLM inference API. Focus on batching requests together before sending them to the GPU. Latency requirement of <500ms-1s, handling high traffic with limited GPUs.
• LLM request batching API: Design an HTTP API where users send individual requests and the system batches them for GPU efficiency.
• Prompt Playground system design: Design a prompt engineering platform like the Anthropic Console. Full-stack design covering product features, UI, architecture, and scalability. Not a chatbot — every run is independent.
• Image processor design: Design a single processor first, then extend to multiple processors. Define APIs for submit, check status, and cancel. Handle concurrency, producer-consumer patterns, and thread safety.
• Prompt sharing system (product design): Not a traditional system design question. Design a prompt-sharing system. Covers user flows, database schema, sharing mechanics, and scaling for viral growth.

Values/Culture Interview

Anthropic asks every candidate about AI safety. A surface-level answer gets filtered out even if your technical rounds were strong.

Reported questions:

• Why does AI safety matter to you personally?
• Do you believe AI will change the world? Would you give up personal benefit to help keep humanity safe?
• Where do you think Anthropic could be doing better on AI safety?
• What are your core values? How do they show up in your day-to-day work?
• Working at Anthropic might mean lower pay than other companies — how do you feel about that?
• Talk about a time you got feedback that hurt, and how you handled it.
• Describe a time you had a conflict with someone and later realized you were wrong.
• How do you handle it when everyone shares the same goal but disagrees on how to get there?
• Why Anthropic? Name one specific value of theirs you agree with, explain why, cite something Anthropic has actually done that reflects that value, and show how you've lived out that value yourself.

How to Prepare

OA: Practice building small standalone programs from scratch. Grinding leetcode alone isn't enough. Specifically practice the multi-stage format — each stage builds on the last, so modular code matters from the very first line.

Technical interviews: Understand how LLM inference actually works in production. Topics worth knowing: request batching, KV cache, GPU memory management, model quantization tradeoffs, and the difference between training and inference infrastructure. Read Anthropic's engineering blog for vocabulary and background.

Values interview: Read Anthropic's core views document. Understand Constitutional AI. Think about what you agree with, what you're unsure about, and what you'd add. They're not looking for agreement — they're looking for evidence that you've actually thought about these questions.

Compensation

Reported Senior SWE offers range $350-500k+ TC, with a large equity component. Given the recent $60B+ valuation, the equity upside is meaningful. Understand the liquidity terms before signing, since it's still a private company.

Hope this helps everyone who's preparing. Good luck.