Anthropic Interview Process: Rounds, Values and Prep

Engineers who prepare for the Anthropic interview process the way they would for any large tech company tend to be caught out twice. The first time is a timed coding assessment that looks nothing like a list of algorithm puzzles and that many people do not finish. The second is an hour of conversation about values with someone who is not an engineer, which candidates and guides alike describe as a round that fails strong technical people.
Neither is a secret, and both can be prepared for. This guide separates what Anthropic itself publishes from what candidates report, goes through the stages for software engineers, and gives the values interview the space it needs. Third-party guides disagree on round counts and timelines; where they do, you get the range, and the advice is to confirm your own format with the recruiter.
What Anthropic publishes
More than most companies, though less than a full process. As of November 2026 the careers page says:
- All interviews are held over Google Meet.
- For technical roles there is live coding "in tools such as Colab and CodeSignal". You may look things up, but you are expected to know basic syntax and the standard library.
- Interviewers also discuss your experience and motivation, and leave time for your questions.
- The company does not give feedback on resumes or interviews.
- A candidate who is turned down can reapply after 12 months, or sooner if their experience has changed materially.
The page also lists the company's values (covered below) and links to a separate page of guidance on AI use by candidates. It does not give the number of rounds, their order or a timeline. All of that comes from candidates.
The stages candidates report
Two of the more detailed 2026 guides, both built on interviews with candidates and engineers, describe the same skeleton and differ on the counts.
| Stage | What is reported |
|---|---|
| Recruiter call | 15 to 30 minutes; motivation and background; sometimes follow-up calls on logistics |
| Coding assessment or screen | 60 to 90 minutes; self-directed on CodeSignal or live in a notebook |
| Hiring manager conversation | 45 to 60 minutes; before the final rounds or part of them |
| Final rounds | Four or five interviews of about an hour, sometimes over two days |
| After the loop | References; team matching in some accounts |
One guide counts three steps and four to five hours of final interviews. The other counts seven or eight evaluative conversations in total. They are describing the same thing with the hiring manager call and recruiter follow-ups counted differently.
Everything is remote. The recruiter call is described as a real screen: candidates are expected to say why Anthropic specifically, as distinct from why AI in general.
The online assessment
For many software engineering candidates the first technical step is a CodeSignal assessment taken alone, with the camera on and no interviewer. Some candidates get a live session in a Python notebook instead, and candidates who come through a referral reportedly sometimes skip the step.
The format, as candidates describe it:
- One problem, several levels. You build something practical, for example the core logic of a bank with several transaction types, and each level adds requirements. Passing the tests at one level opens the next. One engineer described four levels.
- Sixty to ninety minutes. Reports differ, and one account mentions up to two hours.
- Running out of time is common. The task is not conceptually hard. It is a lot of correct code against a clock, with later levels that punish a design that was too rigid at level one.
Preparation that fits:
- Use Python unless you have a strong reason not to, and know the standard library cold:
collections,heapq,itertools,dataclasses, string and date handling. - Practice multi-level builds under a timer: an in-memory database with transactions, a file system with quotas, a scheduler with priorities. Read all visible requirements before typing, then write the simplest structure that will survive new fields and new operations.
- Run the tests early and often. A level you have passed is banked; an elegant refactor at minute 70 is not.
No interviewer means no hints and nobody to hear your reasoning. The score is the tests.
Technical rounds and the policy on AI tools
After the assessment, the reported technical rounds are one or two coding interviews and a system design interview.
Coding. Guides describe practical tasks over puzzles: extend an existing feature, debug code, work out what a partial implementation does, implement a cache and then extend it under new constraints, find duplicate files in a directory tree. Concurrency comes up repeatedly in reports: a multi-threaded crawler, data mutated by more than one caller. The second coding round, where there is one, is said to be specific to the role.
Hiring manager. A walk through your past roles, one by one. Some managers reportedly include a code review exercise.
Project presentation. One guide reports a round in which you present a large recent project you owned end to end, with the format sent in advance by the recruiter.
On AI tools, Anthropic's own guidance page (last updated July 2025, read in November 2026) is specific:
- Application. "Please create your first draft yourself, then use Claude to refine it."
- Take-home assessments. "Complete these without Claude unless we indicate otherwise."
- Live interviews. No AI assistance "unless we indicate otherwise".
So the default is no AI in anything assessed, with exceptions announced by the company. One third-party guide says Anthropic has been piloting an AI-assisted interview that tests whether a candidate can build alongside a model while staying in control of the result. If your recruiter tells you a round allows AI, take that at face value and prepare for it by practicing on unfamiliar code with a strict habit of reading what the tool produced. If they say nothing, assume none, and ask.
The values interview
This is the round candidates prepare for least and guides most often name as the one where people fail.
Anthropic lists seven values on its careers page:
- Act for the global good.
- Hold light and shade.
- Be good to our users.
- Ignite a race to the top on safety.
- Do the simple thing that works.
- Be helpful, honest, and harmless.
- Put the mission first.
The page presents these as how people at the company work together; it does not say they are a scoring rubric. What candidates report is an hour-long conversation, often with an interviewer from outside engineering, made of two kinds of question:
- Reflective. A time your values were tested. A time you pushed back on something. A belief you changed. A technical decision you were wrong about. Follow-ups reportedly go to how it felt and what it cost, as well as what happened.
- Hypothetical. Scenarios without a clean answer. Reported examples include whether you would ship a feature that costs customers more but raises revenue, and whether you have criticisms of the company's direction.
Preparing honestly
The round cannot be passed by memorizing positions, and trying is the common mistake. Guides say enthusiasm scores below considered skepticism, and interviewers are reported to probe an agreeable answer until it either has reasons behind it or collapses.
What does help:
- Read the primary sources. The company's published views on AI safety and its leadership's long-form writing are public. Read enough to know what the company says it is doing and why, and note where you are not convinced.
- Form your own position. On the risk of the technology, on whether a safety-focused lab should build frontier models at all, on what you would not work on. Write half a page. If you have a real criticism, prepare to state it and to say what would change your mind.
- Prepare three true stories. One where you held a position under pressure, one where you changed your mind, one where doing the right thing cost something measurable. Real details beat tidy ones.
- Practice being asked "why" four times. With a person. This is the format of the round, and it is uncomfortable the first time.
If, after the reading, you find you do not care much about the mission, that is worth knowing before the interview and not during it.
System design
Candidates and guides describe this as one of the harder rounds. Two formats are reported: a blank prompt, or a critique of an existing design. One guide says it is run in a shared document, without a drawing tool.
Reported prompts cluster around model serving and its neighbors: a batch inference API, a token-generation service handling on the order of a hundred thousand requests per second, a chat service, a file cache. Themes include request batching and queuing, GPU utilization under variable load, and concurrency. One candidate was reportedly told not to worry about memory limits and to focus on the interesting architectural questions.
For a software engineer without an ML background, the useful preparation is to understand inference as a workload: long and variable request times, streamed output, expensive and limited capacity, batching that trades latency for throughput. The ML and AI systems patterns guide covers the building blocks. The LLM system design interview guide works one such design through from requirements to failure modes, and the AI engineer interview guide lists the fundamentals you may be asked to explain along the way. Because the critique format rewards spotting what is wrong in someone else's design, rehearsing with another person is more valuable here than usual; a generative AI and LLM coach can bring a design for you to take apart and then defend your alternative against.
Timeline and decision
Reported timelines range widely. One guide says three to four weeks in total; another four to six weeks from application to decision, with the final rounds usually scheduled within a couple of weeks of the technical screen; other accounts run to several months when a team pauses hiring.
Decisions are reportedly made by consensus among interviewers, with the hiring manager deciding when there is none. References are said to be taken seriously and to include questions about how you handled conflict. Remember the published policy: no interview feedback, and a 12-month wait to reapply in most cases. That makes it worth being ready before the first call instead of treating the first attempt as practice.
A three-week plan
Week one. Ask the recruiter for the stage list and the AI rules. Do four timed multi-level builds in Python, 90 minutes each, and review where the design blocked you at a later level. Start the reading for the values round.
Week two. Two more builds, with concurrency in each. Three design sessions on inference-adjacent prompts, one of them as a critique of a design you find or a friend writes. Write your half page of positions and your three stories.
Week three. One mock of each: coding with extension, system design, values conversation, hiring manager walk-through of your resume. A coach who works with generative AI and LLMs can run the technical ones. Give the values mock to the most sceptical person you know.
Where coaching fits
techinterview.coach lists no coach with Anthropic on their history, as of November 2026, and this post does not suggest otherwise. What a coach with comparable experience can offer is a timed rehearsal of the multi-level coding format, an LLM system design session in the critique style, and a values conversation with someone who will keep asking why. Four coaches list generative AI and LLMs as a specialty. No coach can tell you what the company's interviewers want to hear, and by the guides' accounts answers prepared that way do badly. For a low-stakes first attempt at talking through your stories, the AI practice interview's behavioral round is free, sign in to start.
Related guides
- OpenAI interview process: the closest comparison, with a different mix of rounds.
- ML and AI systems architectural patterns: serving and batching patterns for the design round.
- Netflix culture interview: another company where a values conversation can decide the outcome.
- Building your story bank: how to prepare true stories that hold up under follow-ups.
- Mock interviews: how to set up a rehearsal that gives usable feedback.
- System design interview questions by level: depth expected at senior and staff levels.