Anthropic Research Engineer mock interview

16-20 day process. CodeSignal take-home with 4 progressive levels — 100% correctness to advance. Values round tests whether you hold beliefs under pressure, not whether you can recite alignment talking points. Performative safety enthusiasm = rejection.

  • ~16-20 days
    End-to-end process
  • 4 levels
    Take-home stages
  • 100%
    Correctness to advance
  • Free
    First mock

Why Anthropic's interview is structured around values pressure-testing, not just coding

Per Anthropic's official careers page (anthropic.com/careers), Greenhouse postings, Sundeep Teki's AI RE guide, and Ridhima Khurana's culture-interview write-up, Anthropic's interview is built to filter on values authenticity, not just technical bar. The values/culture round is non-technical but determinative — many otherwise-strong RE candidates fail here. The round tests whether you hold beliefs under pressure, NOT whether you can recite alignment talking points. Anthropic explicitly wants 'thoughtful disagreement' over agreement, 'shade and light' framing — holding both upside and risk simultaneously.

The CodeSignal take-home is the technical gate. Single problem, 4 progressively harder levels, 100% correctness required to unlock each next level. Partial solutions don't advance. This is unusually strict — most companies grade on partial credit. Per Sundeep Teki, this format filters most candidates before they even reach onsite. Practice timed CodeSignal problems with correctness as the only metric, not speed or elegance.

Compensation tracks the SWE band (no separate RE pay band on Levels.fyi): Senior SWE $563K TC, Lead $785K TC. Equity is 4-year vest with a 25% year-1 cliff, illiquid since Anthropic is private. RE roles do not require an advanced degree per official postings — strong SWE + ML hobby projects + substantive safety engagement (interpretability OSS, alignment forum participation, independent research write-ups) is the documented pathway. Reference checks happen mid-process, distinctively, so have references ready earlier than typical.

What the loop looks like

Round-by-round

  1. 01

    Recruiter call

    Pre-onsite
    30 min

    Background, 'Why Anthropic?', understanding of HHH (Helpful, Honest, Harmless) and AI safety. Performative safety enthusiasm flagged here.

  2. 02

    CodeSignal take-home

    Pre-onsite
    60-90 min

    Single complex problem with 4 progressively harder levels. 100% correctness required to unlock the next level. Distinctive Anthropic format — partial solutions don't advance you.

  3. 03

    Hiring manager

    Onsite
    45-60 min

    Research interests, team direction, fit. 'Why Anthropic specifically' — must reference recent papers / Constitutional AI / interpretability work, not generic safety enthusiasm.

  4. 04

    Coding round

    Onsite
    60 min

    Practical engineering — e.g. 'implement a bank with multiple transaction types' or debug existing code. Python expected.

  5. 05

    ML coding (RE-specific)

    Onsite
    60 min

    PyTorch from scratch. Multi-head attention, training loop debugging, 'loss spikes during pretraining 100B model — debug data, LR, or hardware'. JAX nice-to-have but not required.

  6. 06

    System design

    Onsite
    60 min

    API for serving LLMs efficiently — request batching, queuing, GPU utilization. Distributed training across thousands of accelerators is in-scope.

  7. 07

    Values / culture round

    Onsite
    45-60 min

    Non-technical interviewer. Tests whether you hold values under pressure, NOT whether you recite alignment talking points. Thoughtful disagreement lands better than agreement.

Question bank

Real interview questions reported by candidates

  • Take-home

    "CodeSignal take-home: single complex problem with 4 progressive levels. Must achieve 100% correctness on each level to unlock the next. Practical coding (data manipulation, system design in code)."

    Source · sundeepteki.org AI RE interview guide 2026, igotanoffer.com Anthropic interview process
  • ML coding

    "Implement Multi-Head Attention from scratch in PyTorch — masking, scaled dot-product, multi-head split/merge. Anthropic baseline."

    Source · interviewing.io Anthropic, sundeepteki.org AI RE interview guide
  • ML debugging

    "Debug a broken training loop. Anthropic-flavored: 'During pretraining of a 100B model you see loss spikes. How do you debug — data, learning rate, or hardware?' Walk through systematic diagnosis."

    Source · sundeepteki.org AI RE interview guide 2026, linkjob.ai Anthropic
  • Coding

    "Implement a bank with multiple transaction types — practical engineering test, not LeetCode. Tests clean code, edge cases, testing strategy."

    Source · interviewing.io Anthropic
  • System design

    "Design an API for serving LLMs efficiently. Cover request batching, queuing strategy, GPU utilization, latency vs throughput trade-offs, fault tolerance."

    Source · interviewing.io Anthropic, anthropic.com greenhouse RE postings
  • Research discussion

    "Discuss a recent Anthropic paper (Constitutional AI, interpretability work). What did you find compelling? What concerns you? What would you push back on?"

    Source · anthropic.com careers, ridhimakhurana.substack.com Anthropic culture interview
  • Behavioral / values

    "Tell me about a time you held an unpopular technical opinion. How did you handle it? What did you change your mind on?"

    Source · ridhimakhurana.substack.com Anthropic culture interview
  • Values

    "What are the strongest arguments AGAINST Anthropic's approach to AI safety? Where do you disagree with the constitutional AI framework?"

    Source · anthropic.com careers, ridhimakhurana.substack.com
Common signals to fix

What gets you rejected at this level

  • Performative AI-safety enthusiasm

    Saying 'I'm deeply aligned with your mission' or generic 'AI could be dangerous' without specifics reads as inauthentic. Anthropic explicitly does NOT want recited alignment talking points. Per Ridhima Khurana's culture-interview write-up: 'thoughtful disagreement lands better than agreement.' Be ready to articulate where you'd push back on Constitutional AI or RLHF.

  • Dismissiveness toward safety

    On the other end — being dismissive of safety concerns, treating them as PR theater, or showing impatience with values discussions is a major flag (per Sundeep Teki). The values round is non-technical but determinative. Many otherwise-strong candidates fail here.

  • Take-home didn't reach 100% correctness

    Anthropic's CodeSignal take-home has 4 progressive levels with 100% correctness required to unlock the next. Per Sundeep Teki, this is the gate that filters most candidates. Partial solutions don't advance. Practice timed CodeSignal-style problems with correctness as the only metric.

  • Can't write efficient PyTorch under time pressure

    Tensor-shape fluency, contiguity, masking, view ops are common failure modes (per linkjob.ai aggregation). 'I know transformers in theory' = fail if you can't implement attention from scratch in 30 min. Practice from-scratch implementation before the loop.

  • Over-engineering the take-home

    Anthropic wants 'working code, clear reasoning, practical trade-offs' — not algorithmic elegance or framework abstraction. Candidates who add unnecessary class hierarchies, premature optimization, or over-abstracted designs flag missing pragmatic engineering judgment.

  • No personal-stakes story

    'Why do you care about safe AI?' — generic answers fail. Anthropic wants candidates with substantive engagement: independent safety blog posts, OSS contributions to interpretability, alignment forum participation, or specific framing of why your background led you here. Without it, mission alignment reads hollow.

How Applr's AI mock interview tracks Anthropic's RE rubric

Applr's Anthropic RE mock simulates the progressive CodeSignal take-home format — 4 levels with 100% correctness required to advance, no partial credit. ML coding rounds prompt PyTorch from-scratch implementation (multi-head attention, training loop debugging, 100B-model loss-spike diagnosis).

The values round prompts thoughtful-disagreement scenarios — performative 'I love your mission' answers get flagged for rewrites that surface substantive engagement with where Constitutional AI or RLHF could go wrong. Personal-stakes framing for why you care about safe AI gets prompted explicitly; without it, mission alignment reads hollow.

Adjacent loops

Related practice

FAQ

Do I need a PhD for Anthropic Research Engineer?

No, per Anthropic's Greenhouse postings and official careers page. RE requires Bachelor's + significant SWE experience, Python, deep-learning framework proficiency. Advanced degree NOT required. Anthropic publicly states 'about half of technical staff have PhDs… plenty of brilliant colleagues never went to college' (anthropic.com/careers). For international students specifically: Anthropic rarely hires straight from undergrad without exceptional documented contributions to AI research or open-source safety projects (per letsdatascience.com). Strongest signals are interpretability/safety OSS contributions, independent research write-ups, or production ML infra experience — NOT GPA or competition results.

What's the actual salary at Anthropic?

Per Levels.fyi (proxy via SWE since no separate RE band): Senior SWE $563K TC ($316K base + ~$247K equity), Lead SWE $785K TC ($332K base + $453K equity). Median across Anthropic: $420-710K depending on level. Research Scientist median $746K, top $1.05M. UK Science of Scaling RE band: £260K-£630K (per Anthropic Greenhouse). Equity is 4-yr vest, 25% year-1 cliff, monthly thereafter — illiquid (Anthropic is private). Liquidity historically via tender offers, but frequency/terms not publicly confirmed.

What is the take-home actually like?

Per Sundeep Teki and IGotAnOffer Anthropic write-ups: CodeSignal-hosted, 60-90 min, single complex problem with 4 progressively harder levels. CRITICAL: 100% correctness required to advance to the next level. Partial solutions don't unlock anything. Most candidates fail here. The problem is practical (not pure algorithms) — data manipulation, system design in code, or ML-flavored implementation. Practice timed CodeSignal problems with correctness as the only goal — not speed, not elegance.

How is the values round graded?

Per Ridhima Khurana's culture-interview write-up (ridhimakhurana.substack.com/p/inside-anthropics-culture-interview): non-technical interviewer; tests whether you HOLD values under pressure, not whether you can recite them. Anthropic explicitly wants 'thoughtful disagreement' over agreement. 'Shade and light' framing — hold both upside AND risk simultaneously. Performative 'I love your mission' = fail. Dismissive 'safety is PR' = fail. Substantive engagement with where Constitutional AI or RLHF could go wrong, with personal-stakes story for why you care, lands.

Should I learn JAX before applying?

Per the Anthropic RE Science of Scaling Greenhouse posting: PyTorch is REQUIRED, JAX is 'nice to have'. If you're choosing where to invest time before the loop, prioritize PyTorch fluency (from-scratch transformer implementation, training loop debugging, distributed training basics). JAX is a plus but won't make up for shaky PyTorch.

Reference checks happen WHEN?

Mid-process, NOT at offer stage. This is distinctive to Anthropic per Sundeep Teki. Have references aware and ready earlier than typical FAANG processes. Don't tell references they're being contacted only at offer time — they may be contacted before your final loop.

Ready to practice for Anthropic?

Start free →