Anthropic Interview Guide
Anthropic has a 6-question recurring bank, a code review HM round, and values screening that fails more candidates than the technical rounds. The company that builds Claude redesigns its own interview questions when Claude beats them. If you can articulate a genuine perspective on AI safety, you have a structural advantage.
What makes Anthropic different
Anthropic was founded as a safety-conscious fork of OpenAI. Dario and Daniela Amodei left in 2021 over philosophical differences about building safety into models from the beginning. That origin story permeates every hiring decision: the most common reason candidates fail is the culture fit round, not the technical rounds. If you cannot articulate a thoughtful perspective on AI risks, you will not advance regardless of technical ability.
The coding interview uses a tiny, known bank of 6 problems that recycle across candidates: Web Crawler, LRU Cache, Stack Trace, Distributed Mode/Median, Profiler Trace, and Tokenization. These are progressive multi-level problems on CodeSignal where you pass each level's unit tests to unlock the next. With only 6 questions, most candidates know them beforehand — but interviewers drill into concurrency, distributed extensions, and failure modes that pure memorization cannot cover.
Two things stand out about Anthropic's interview. First, the hiring manager round is code review, not coding — you analyze someone else's codebase to find bugs and concurrency risks. Second, Anthropic's engineering team publishes openly about designing “AI-resistant” evaluations and iteratively rebuilds questions when Claude models outperform candidates. When Opus 4 beat their performance engineering take-home, they created v2. When Opus 4.5 matched top human performance, they shifted to Zachtronics-style puzzle problems. The interview format evolves faster here than at any other company.
Everyone carries the title “Member of Technical Staff” — from new hires to co-founders. Internally it's T4/T5/T6+, but LinkedIn cannot differentiate. About 50% of technical staff had no prior ML experience, and about 50% hold PhDs. The company has surged from ~500 employees in late 2023 to ~5,000 in 2026, with pre-IPO equity at a $380B valuation making the comp package among the most equity-leveraged in tech.
The interview loop
6–8 total interviews across 4–6 stages. Two recruiter calls, coding screen, HM code review, 4-round onsite. Consensus-based hiring with HM final authority.
Recruiter Screens (2x)
30 min each · Phone / VideoFirst call covers background and motivation. Second call discusses compensation AND distributes culture/values study documents. Read those documents — the values round is where most candidates fail.
Technical Coding Screen
60–90 min · CodeSignalgateOne progressive multi-level problem with escalating complexity. Must pass each level’s unit tests to unlock the next. Six known recurring problems. Google/Stack Overflow allowed; AI tools are not permitted unless Anthropic says otherwise.
Hiring Manager Interview
60 min · Code ReviewgateCode review, not coding. You analyze existing codebases to find bugs, concurrency risks, and architectural issues.
Onsite: System Design
60 min · Whiteboard / VirtualLLM-infrastructure-native: inference batching, GPU scheduling, token-generation services, distributed search. Not "Design Instagram." Evaluation scales with level: junior = clarity, senior = trade-offs + failure scenarios.
Onsite: Coding
60 min · Your IDE or CodeSignalgateReal-world progressive problems with escalating parameters. Practical engineering focus — modular, extensible code that adapts to new requirements. Python strongly favored.
Onsite: Experiences & Goals
60 min · BehavioralManager-level interviewer probing collaboration, conflict resolution, candor. Values screening is woven throughout. AI ethics and safety questions are not a checkbox — they are a hard gate.
Onsite: Project Presentation
60 min · PresentationPresent a past project showing end-to-end ownership. Interviewers probe organizational impact and tradeoff rationale more than technical depth. Consensus-based hiring with HM final authority.
The code review round — what you actually need to know
The hiring manager round is code review, not code writing. You'll analyze an existing codebase to identify bugs, concurrency risks, and architectural bottlenecks. Airbnb's code review round is the closest analog.
How to prepare: Practice reading unfamiliar Python codebases. Look for threading issues (race conditions, deadlocks, GIL implications), error handling gaps, and architectural anti-patterns. The focus is on identifying problems in real code, not solving greenfield puzzles.
This section is in Talk with Coco: $79 once for 3 months of hosted voice.
See pricingValues screening — the Anthropic-distinctive section
This is what makes Anthropic interviews fundamentally different from every other company, including OpenAI. Seven core values are explicitly assessed:
- Act for the global good
- Hold light and shade (acknowledge both risks and benefits)
- Be good to users, policymakers, and affected communities
- Ignite a race to the top on safety
- Do the simple thing that works
- Be helpful, honest, and harmless
- Put the mission first
Behavioral red flags that directly disqualify: lone wolf mentality, arrogance, purely financial motivation, inability to discuss AI ethics beyond surface-level boilerplate, and discomfort with candid discussion of past failures. The recruiter distributes culture documents early — treat them like a take-home assignment.
New grad entry (MTS)
New grads enter as Member of Technical Staff (MTS) — the same external title everyone carries. Internally likely T3 or T4. H-1B filings show MTS base range $300K–$405K. Zero bonuses. 4-year vesting with pre-IPO RSU-equivalent equity at $380B valuation.
What's different for new grads:
- The 6-problem recurring bank applies to all levels. Studying it has high ROI.
- Values screening is the primary failure point. Read the culture documents the recruiter provides. Study Anthropic's safety publications. “I want to work on cool AI” is not enough.
- The HM code review round tests reading, not writing. Practice spotting bugs in unfamiliar Python.
- ~50% of technical staff had no prior ML experience. If you're a new grad without ML background, that's fine.
- 12-month cooldown after rejection. Longer than Meta (3–6 months).
- Hybrid Bay Area with monthly visits. Remote employees report feeling “out of the loop.”
SWE vs Research Scientist tracks
Anthropic organizes interviews by two orgs: Applied (SWE) and Research. Your interviewers all come from your target org. Key differences:
| Dimension | SWE (Applied) | Research Scientist |
|---|---|---|
| Coding focus | Practical systems, progressive complexity | Algorithm reasoning + collaborative problem-solving |
| Design round | LLM infrastructure, distributed systems | ML theory + system architecture |
| Unique round | Code review of existing codebases | Research presentation / case study |
| Glassdoor | 29% positive | 100% positive |
| Prep | Study team-specific infra | Review 40+ Anthropic papers |
Interview culture
Candidates consistently describe Anthropic's interview as “thoughtful” and less adversarial than FAANG. One interviewing.io respondent noted the process “felt so easy and thoughtful compared to all the other companies I interviewed with. They have their shit together.”
However, the 29% positive rate for SWE interviews on Glassdoor suggests significant variability in execution — possibly reflecting the rapid scaling from ~500 to ~5,000 employees straining interview calibration. Research Scientists report 100% positive, suggesting the research loop is more refined.
The emphasis on practical engineering over algorithmic tricks, the code review round, and the genuine (not performative) values discussion create a distinctive interview experience. CEO approval is 93%, recommend-to-friend is 95%, and comp & benefits rated 4.8/5.0 (highest category). Work-life balance is 3.7/5.0 — some 60+ hour weeks are reported.