Company interview guide

Anthropic Machine Learning Interview Guide (2026)

A timed, level-by-level build exercise, practical live coding, LLM infrastructure design and a values interview that really counts.

Anthropic's machine learning interview loop has 9 reported stages, from online assessment to references and team match; typical length: 3-8+ weeks.

Roles covered: ML Engineer, Research Scientist and AI Engineer. The question bank holds 50 questions reported from Anthropic interviews, 5 of them free to solve in your browser.

Try a free question

Interview process length
3-8+ weeks
Stages
9
Reports linked to bank questions
59
Bank report dates
Dec 2023–Sep 2026
Questions in bank
50 (5 free)
Last reviewed
October 2026

How the Anthropic interview loop works

Most engineering and research candidates start with a timed online assessment, then have a recruiter call (some add a hiring-manager call), a live technical screen with an engineer, and a virtual onsite of four or five rounds. Reported onsites combine practical coding or debugging, a system design round, a role-specific ML round for some teams, a project presentation or deep dive, and a culture interview. References are checked carefully and have ended loops. Performance-engineering and fellowship tracks have used multi-hour take-homes instead of, or on top of, these steps. Gaps of two to three weeks between steps are common, and the company does not give feedback after a rejection.

The Anthropic interview process, stage by stage

01Online assessment

90 min on CodeSignal in Python: one system built over 4 levels against visible unit tests, no AI tools

The most consistently reported step across ML engineer, research engineer and research scientist applicants, and usually the first one; a referral has let at least one candidate skip it. You implement a small in-memory system and each level adds rules that force you to extend or refactor what you wrote. Banking systems come up most often; in-memory databases, file stores and time-tracking systems also appear. A fellowship candidate reported a proctored version plus a second 60-minute round fixing failing tests in an existing Python codebase.

What it tests

  • Turning a long written spec into correct code quickly
  • Data modelling that survives new requirements at levels 3 and 4
  • Timestamps, ordering, expiry (TTL) and history queries
  • Ranking and aggregation, such as top accounts by outgoing money
  • Keeping earlier tests green while adding features

How to prepare

  • Skim all four levels before writing code so your data model already fits the merge or history level.
  • Budget roughly 20-25 minutes per level and move on once the tests pass; running out of time on the last level is the usual failure.
  • Practise in a browser editor without AI help; several candidates were rejected with near-perfect scores, so clean code matters too.
  • If you get a debugging round, start from the failing test and fix the root cause instead of patching for one test.

A 90-minute levelled CodeSignal build in 19 candidate reports from 2022 to 2026, mostly research and ML roles; official pages describe take-home technical assessments but not this levelled format.

02Recruiter and hiring-manager calls

Recruiter call, sometimes followed by a hiring-manager call; lengths not stated in reports

The recruiter checks background, motivation and level. Some candidates then have a hiring-manager call before the onsite; a data-science candidate was asked to design a metric there, and research candidates report questions on their research background and fit. Several research engineers got the online assessment before speaking to anyone, so this call may come after it.

What it tests

  • A specific answer to why Anthropic, and why this role
  • How your past projects map to the team: decisions, trade-offs, results
  • Logistics: level, location, timeline

How to prepare

  • Ask what each later round covers; screens arrive with only a short hint text, not a problem name.
  • Prepare two or three projects you can defend in depth, including what you would change.
  • Have a specific answer to why Anthropic ready from the first call; it is asked throughout the loop.

Recruiter call in 8 candidate reports, five of them non-ML roles (counted as a shared stage); 3 add a hiring-manager call before the onsite. Several engineers took the assessment first.

03Live technical screen

About an hour with an engineer in a shared editor, notebook or Colab; a short topic hint is emailed in advance

Usually a practical, multi-part coding problem where the interviewer adds requirements as you go, such as turning profiler stack samples into trace events, tracking cluster status updates, or reasoning about a MapReduce-style job and how to parallelize it. One ML-track screen asked for attention-free transformer code in Colab, and one was an inference-serving design discussion instead of coding.

What it tests

  • Understanding an unfamiliar spec quickly and asking the right clarifying questions
  • Clean state handling for event streams and nested structures
  • Reasoning about parallel computation and network cost
  • Explaining your plan while you code

How to prepare

  • Two candidates spent about 30 minutes just understanding the prompt; restate the problem with a small example before coding.
  • Talk through the approach for follow-ups you cannot finish; one candidate passed after only explaining the second part.
  • Use the hint text to drill the matching topic, for example map-reduce patterns when it mentions distributed efficiency.
  • SQL items here come from public topic lists for data roles; skip them unless your role touches analytics.

Free reported questions for this stage

Live screen with an engineer in 10 ML, research and staff-engineer reports since 2022, mostly practical multi-part coding; the careers page says technical roles use live coding in Colab and CodeSignal.

04Onsite: coding and debugging

One or more live rounds in Colab or CodeSignal; some loops use a Python debugging exercise instead of writing from scratch

Onsite coding stays practical: reports describe applied problems related to the role and a round that hands you an existing Python codebase with failing behaviour to fix. Expect the same style as the live screen, with growing requirements, concurrency and state handling rather than puzzle algorithms. One candidate was switched from Colab to CodeSignal mid-loop.

What it tests

  • Writing correct, readable code for a practical spec under time pressure
  • Reading someone else's code and finding the root cause of failing tests
  • Concurrency: async workers, limits, retries and shared state
  • Extending a design when the interviewer adds a requirement

How to prepare

  • Practise debugging unfamiliar repos against their tests, not only writing fresh code.
  • Be comfortable with Python threading and asyncio; batching and rate-limit problems rely on them.
  • Try both Colab and the CodeSignal editor beforehand so a tool switch does not cost you time.

3 reports: one coding round in a staff-engineer onsite, three unspecified technical rounds in an ML loop, a Python debugging round in an AI/RL loop; the careers page confirms live coding in Colab and CodeSignal.

05Onsite: system design

Open-ended discussion, often with numbers to estimate; length not stated in reports

Reported prompts are mostly about running large models: scheduling large and small model batches on one GPU pool, routing traffic across mixed inference backends, and performance modelling of a low-latency multi-region service. Compiled summaries of several candidates add shipping a 500 GB model to a GPU fleet and a batched inference API. Follow-ups change the constraints mid-way.

What it tests

  • Back-of-envelope throughput, bandwidth and latency estimates
  • Batching, queueing and scheduling with fairness and tail-latency targets
  • Failure handling, retries and safe rollout of new model versions
  • Caching and when it is valid (same model version, deterministic settings)
  • Adapting the design when a new constraint arrives

How to prepare

  • Know LLM serving basics: prefill vs decode, KV cache memory, continuous batching, tensor parallelism.
  • Work the numbers out loud, for example transfer time for 500 GB over 10 Gbps, before picking an architecture.
  • Expect follow-ups that change the shape of the problem (10x more workers, a second model type) and re-plan calmly.

Free reported questions for this stage

2 reports for covered roles (an onsite design round, a design-style screen on GPU scheduling); two manager loops and compiled multi-candidate summaries add LLM-serving prompts.

06Onsite: ML and LLM rounds (role-specific)

Depends on the team: PyTorch debugging in Colab or CodeSignal, tensor coding, or a role-related discussion

The content depends on the team and is thinly reported. First-hand reports mention a PyTorch debugging round and a discussion of applied problems related to the role; an ML-track screen asked for an attention-free transformer from scratch. Public topic lists add transformer internals, scaling laws, Constitutional AI and evaluation of LLM systems; treat those as likely rather than confirmed.

What it tests

  • Tensor shapes and matrix manipulation for attention and its variants
  • Finding bugs in model or training code someone else wrote
  • Transformer and serving internals: masks, KV cache, grouped-query attention
  • Alignment methods such as RLHF and Constitutional AI, and their limits
  • Designing evaluations for LLM behaviour, hallucination and regressions

How to prepare

  • Be able to write causal multi-head attention in numpy or PyTorch from memory, then adapt it to a variant you have not seen.
  • Prep material may not match the round exactly; focus on fluent matrix reasoning rather than memorized code.
  • Read Anthropic's papers relevant to your team so you can discuss methods like Constitutional AI with real opinions.

Free reported questions for this stage

2 reports: a role-related discussion round in an ML engineer loop and a PyTorch debugging round in an AI/RL loop; the topic list beyond that comes from public write-ups.

07Project deep dive or presentation

A technical presentation on past work or rounds probing your projects; public write-ups describe a short slide talk plus questions

You walk through work you owned while interviewers probe decisions, trade-offs and what went wrong. An AI/RL candidate gave a technical presentation; an ML engineer had several rounds on past ML work and problem solving. Manager loops call it a project retrospective and report sharp follow-up questions.

What it tests

  • Ownership and depth behind the claims on your resume
  • Trade-offs you made and how you knew they worked
  • Working across teams and handling setbacks

How to prepare

  • Pick one project with real technical depth and rehearse the follow-ups: scale, alternatives, your exact contribution.
  • Have a clear story for a failure and what you learned from it.

2 reports for covered roles (a technical presentation; several rounds on past ML work); both manager loops had a project retrospective. The company's fellowship posting lists a research discussion.

08Culture and values interview

Conversational round on values and AI safety, often with a shadow interviewer; length not stated in reports

Candidates across roles describe a dedicated culture interview about AI safety and Anthropic's values. Two manager candidates with positive technical feedback were turned down on culture fit, one after references, so treat it as a real gate.

What it tests

  • Why AI safety matters to you personally, and why Anthropic over other labs
  • How a specific company value shows up in your own past work
  • Handling hard feedback, being wrong in a disagreement, and conflict
  • Thinking about risks and trade-offs of the technology honestly

How to prepare

  • Read Anthropic's Core Views on AI Safety and its listed values; form your own view, including what you would question.
  • Prepare a respectful answer to why Anthropic rather than a competitor; a candidate was caught off guard by it.
  • Use concrete stories, not slogans, and know which company value matches your own past work.

Culture or values round in 8 candidate reports across ML, staff-engineer, data and manager roles (shared stage); the careers page says interviews ask what motivates you, but does not name a values round.

09References and team match

Reference calls after the onsite; team matching described only in public write-ups

References are contacted after the onsite and are not a formality: one hired candidate stressed how seriously they are taken, and a manager candidate was rejected two days after reference checks. Public write-ups add a team-matching step, which our reports do not describe.

What it tests

  • Consistent evidence of impact and collaboration from people who worked with you
  • Fit with an open team

How to prepare

  • Choose references who can describe specific work, and brief them in advance.
  • Follow up politely if you hear nothing for about 10 days.

2 reports call reference checks decisive. The company's fellowship posting lists a reference check (at the application stage, earlier than reports describe) and a mentor matching step.

Questions and themes reported most often

  • Four-level banking system: transfers, top spenders, scheduled payments, merged accounts (5 reports, online assessment)
  • Why Anthropic, and why AI safety matters to you (4 reports, culture and values interview)
  • Presentation or deep dive on a project you owned (4 reports, project deep dive or presentation)
  • Distributed or parallel computation in the live screen (map-reduce, parallelize a task) (2 reports, live technical screen)
  • Fix failing tests in an existing Python or PyTorch codebase (2 reports, onsite: coding and debugging)
  • LLM inference scheduling, batching and routing across a GPU pool (2 reports, onsite: system design)
  • Time-tracking database for employees or factory workers (1 reports, online assessment)
  • Convert sampled profiler stacks into start and end events (1 reports, live technical screen)

What Anthropic asks about

  • Practical & progressive coding45%
  • LLM infrastructure system design20%
  • Culture, values & deep dive20%
  • ML & LLM coding and concepts15%

Anthropic interview prep checklist

  • Finished a four-level banking or in-memory database build inside 90 minutes with all visible tests passing
  • Solved the stack-samples-to-events problem and a map-reduce efficiency question without hints
  • Fixed failing tests in an unfamiliar Python or PyTorch codebase within an hour
  • Can design GPU scheduling and batching for an inference service with numbers
  • Wrote causal multi-head attention from memory and adapted it to a variant
  • Have a personal, specific answer to why Anthropic and why AI safety
  • Prepared a project presentation, including a failure and what you changed
  • Chosen and briefed references who can speak to concrete impact

Tips

  • Treat the online assessment as the main gate: drill banking, key-value and file-store builds over four levels in 90 minutes.
  • Read every level before you start and design for the last one; refactoring at level 3 under time pressure is the usual failure.
  • Practise without AI tools; the company's guidance bans AI in assessments and live interviews unless you are told otherwise.
  • For live coding, clarify the spec with a tiny example first and keep talking through your plan.
  • Practise debugging an unfamiliar Python or PyTorch codebase against failing tests.
  • Learn LLM serving numbers well enough to estimate bandwidth, KV cache memory and batch throughput on a whiteboard.
  • Prepare the values interview as seriously as the technical rounds, with your own considered view on AI safety.
  • Pick references early and brief them; this step has decided outcomes.

See the rest of the Anthropic loop

The later stages, the questions reported for them and how to prepare are in Pro.

Unlock with Pro

Practice the Anthropic loop

Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.

Anthropic interview FAQ

How many interview rounds does Anthropic have for ML roles?

Candidates report 9 stages: Online assessment, Recruiter and hiring-manager calls, Live technical screen, Onsite: coding and debugging, Onsite: system design, Onsite: ML and LLM rounds (role-specific), Project deep dive or presentation, Culture and values interview and References and team match. Most engineering and research candidates start with a timed online assessment, then have a recruiter call (some add a hiring-manager call), a live technical screen with an engineer, and a virtual onsite of four or five rounds. Reported onsites combine practical coding or debugging, a system design round, a role-specific ML round for some teams, a project presentation or deep dive, and a culture interview. References are checked carefully and have ended loops. Performance-engineering and fellowship tracks have used multi-hour take-homes instead of, or on top of, these steps. Gaps of two to three weeks between steps are common, and the company does not give feedback after a rejection.

How long does the Anthropic ML interview process take?

Reported timelines run 3-8+ weeks (3 weeks to 2+ months).

What does Anthropic ask in ML interviews?

By reported question mix: Practical & progressive coding 45%; LLM infrastructure system design 20%; Culture, values & deep dive 20%; ML & LLM coding and concepts 15%. The most repeated themes: Four-level banking system: transfers, top spenders, scheduled payments, merged accounts; Why Anthropic, and why AI safety matters to you; Presentation or deep dive on a project you owned.

How should I prepare for the Anthropic ML interview?

Treat the online assessment as the main gate: drill banking, key-value and file-store builds over four levels in 90 minutes. Read every level before you start and design for the last one; refactoring at level 3 under time pressure is the usual failure. Practise without AI tools; the company's guidance bans AI in assessments and live interviews unless you are told otherwise.

Are there free Anthropic interview questions?

Yes. LeetML has 50 questions reported from Anthropic interviews, and 5 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources