Anthropic Interview Process 2026: Rounds and How to Prepare

If something here is out of date, please email hello@tryorthic.com with what has changed and, if you can, where you saw it.

Short answer

Anthropic's careers page describes how its interviews run but doesn't list the rounds. Most accounts come from software engineers who interviewed in 2025 and 2026, and they usually describe these rounds, in this order:

1. A recruiter call.
2. A timed coding assessment, often called the online assessment.
3. A live coding screen.
4. A conversation with the hiring manager. Some candidates have it as part of the onsite instead.
5. A virtual onsite of four or five one-hour interviews.
6. References and team matching, where you meet teams that might hire you.

Orthic, which publishes this guide, isn't affiliated with Anthropic.

What Anthropic says about its interviews

This section draws on four things Anthropic has published: its careers page, its guidance for candidates on using AI, an engineering blog post about one team's take-home assessment, and its job posts.

Where and how. All interviews run on Google Meet. Technical interviews use live coding tools such as Colab and CodeSignal. You may look things up while you code, though Anthropic doesn't list what you may use, and it expects you to be comfortable with basic syntax and your language's standard library. Interviews for non-technical roles are conversations.

Using AI. Anthropic's guidance for candidates, "How to collaborate with Claude during our hiring process" (last updated July 2025), sets these rules:

• You write the first draft of your application yourself. You may then use an AI tool such as Claude to refine the wording, but not to describe experience you don't have.
• You may use AI to prepare. Anthropic encourages using Claude to research the company, practise answers and prepare questions.
• You complete take-home assessments without AI, unless Anthropic says otherwise for that assessment.
• You answer live interviews without AI help, unless your interviewer says otherwise.

The guidance doesn't mention the timed coding assessment by name. Our advice is to assume no AI unless the instructions for it say otherwise.

In May 2025, Anthropic asked applicants not to use AI in their applications. Press reports in July 2025 said it had replaced that request with these stage-by-stage rules. Some older advice still describes the earlier request.

Anthropic's performance engineering team does allow AI in its take-home assessment. In a January 2026 blog post, the team explained that candidates have been given slow code for a simulated processor since early 2024 and asked to make it run in fewer processor cycles, first in four hours and later in two. More than 1,000 candidates have taken it. Each new Claude model forced the team to redesign the task. The latest version is a set of small optimisation problems scored by how few instructions a solution uses.

After the interviews. Anthropic doesn't give feedback on applications or interviews. If you aren't successful, you can apply again after 12 months, or sooner if your experience has changed significantly.

Background. Anthropic says about half of its technical staff had no machine learning experience before joining. It asks engineers who are interested in research to apply as engineers.

What hiring managers look for now. Anthropic's hiring managers now write more about a technical candidate's autonomy (working out what to do without being told) and judgment (making sound calls when the right answer isn't obvious), and less about expertise in one specific field, according to research Anthropic published in October 2026. In 2024, about 5% of their write-ups on technical candidates mentioned the candidate's record with AI tools; by 2026 that share had tripled. Junior staff are also a much smaller share of new technical hires: 27% in early 2025, and 9% by the time of the report.

Scams. Anthropic's recruiters only email candidates from @anthropic.com addresses, and Anthropic never asks for money or bank details before your first day.

The rounds, step by step

This picture comes from engineers' public accounts of their interviews and from press coverage.

How long it takes. Usually three to six weeks. One review site's average across all 209 of its Anthropic reports, for every role, was 24 days when we checked. Ask your recruiter for your own timeline.

What each interview tests

Coding. Candidates describe the timed assessment, the live screen and the onsite coding interviews as practical tasks rather than puzzle-style algorithm questions: build a small working program, then change it as new requirements arrive, sometimes ending with making it thread-safe, so that several threads can use it at once without corrupting its data. One candidate was told to expect no puzzle-style algorithm questions.

System design. Expect questions on API design, data storage, scalability and fault tolerance. Several engineers were also asked about inference, which means running trained models to answer requests. These cover batching (grouping several requests into one call to the model so the GPU is used efficiently), scheduling (deciding which GPU handles each batch), and what the service does when it is overloaded, such as rejecting some requests (load shedding) or making senders slow down (backpressure). In 2026, three candidates said they reviewed an existing, partly finished design document for an inference service instead of designing from a blank page. They pointed out what was missing and answered questions about how it would fail under heavy load.

Culture interview. Press reports from August 2026 say this interview is run by an employee nominated for the task, that its questions are suggested rather than scripted, and that candidates are asked about a moral dilemma they faced. Engineers also describe follow-up questions about how they felt at the time and what they would do differently now. Press reports from May 2026 say the interview is part of the onsite for nearly every role, and that recruiters and coaches say technically strong candidates are often rejected there.

Hiring manager and project presentation. The hiring manager tends to ask why you made your decisions, what you gave up, what went wrong and what you would change. In the project presentation, candidates for senior and staff roles present a project they led from start to finish: why it existed, its phases, the technology, the team and how it shipped. The interviewers then ask about any part of it. One candidate was asked how the project would scale to several regions (a multi-region design), how its data would be split across machines (sharding) and how it would stay reliable.

Anthropic interview questions

Candidates report being asked the questions below. Under each question, "sources" counts the sites or publications that reported it and "reports" counts individual candidates' accounts. Anthropic hasn't confirmed any of them, and questions vary between interviewers.

We don't publish the timed coding assessment's stage-by-stage requirements, because that detail would help people cheat on it.

Applied AI Engineer and Forward Deployed Engineer

Both roles build products with Claude for Anthropic's business customers. An Applied AI Engineer advises customers on how to design their applications, builds evaluations (sets of test cases that check a model's answers) and prototypes, and runs workshops. A Forward Deployed Engineer works inside a customer's organisation and builds production applications with them. Anthropic's current job posts for these roles ask for:

• Four or more years in a technical role. The Forward Deployed Engineer posts also ask for customer-facing experience.
• Production experience with large language models: prompt engineering (writing and testing the instructions given to a model), building agents, evaluations, and MCP, the Model Context Protocol, a standard way to connect a model to tools (functions it can call) and data.
• Python. The Applied AI Engineer post also accepts TypeScript.

Anthropic also advertises a staff-level Forward Deployed Engineer role. Its post asks for experience designing secure production systems inside large enterprises, where a design has to work with older systems, with the company's identity and access management (who can sign in and what each person can reach) and with compliance requirements. It also expects hands-on experience building with large language models and agents.

Very little is public about the interviews for these roles. One site that covers them describes a role-play in which a customer wants Claude to answer questions from their company's documents, which is usually built with retrieval-augmented generation (RAG), and a design interview on how Claude would be deployed for a customer. It says its own information is thin. We found no report from a candidate who went through these interviews, so ask your recruiter which rounds you'll have and whether one is a customer role-play.

This guide doesn't cover research roles.

How to prepare, round by round

Recruiter call. Prepare an answer to "Why Anthropic?" that names something specific: a piece of its published research, its writing on AI safety, or its Responsible Scaling Policy (Anthropic's published rules for which safeguards must be in place before it trains or releases more capable models). Say which part you agree with and which part you'd question.

Coding interviews. Set a 30-minute timer and build a small program, for example an in-memory key-value store. Then give yourself two new requirements, such as values that expire after a set time (a TTL, or time-to-live) and making it thread-safe, and add them without rewriting what you had. In your main language, practise starting threads, protecting shared data with a lock, using a thread pool (a fixed set of threads that take work from a queue) and a bounded queue (one that blocks or rejects new work when it is full).

System design. For each design choice, practise saying out loud what you chose, what you could have chosen instead, and what your choice cost, such as latency, money or complexity. For inference questions, be ready to explain when you'd send a batch (when it is full, or after a time limit), how you'd schedule batches across GPUs, for example by sending each batch to the GPU with the shortest queue, which requests the service rejects first when it sheds load and what it tells senders when it applies backpressure, and which metrics you'd monitor to know it was failing, such as queue length, latency and error rate.

To practise this, Architecture Workshop gives you AI system design problems to answer stage by stage, as in a real interview, with written feedback on each answer. For questions about what happens when things go wrong, Incident Room puts you in charge of an AI system that has failed in production: you work through its logs, metrics and deploy history to find the cause, then decide how to fix it.

If you're applying for an Applied AI or Forward Deployed Engineer role, where customers often want Claude to answer from their own documents, our free articles on how retrieval-augmented generation works and finding the right information explain retrieval: how an application searches a set of documents for the passages that answer a question and adds them to the model's prompt. To design one yourself, try Design a Secure Multi-Tenant RAG System in Architecture Workshop: an assistant that answers employees' questions from company documents, where many companies (tenants) share one system and nobody may see a document they aren't allowed to read. To see how such a system fails, Investigate a Cross-Tenant RAG Data Leak in Incident Room starts with an employee receiving an answer drawn from another company's private document: you find out how it happened, stop the leak and decide when it's safe to restart.

Culture interview. Prepare examples from your own work: a disagreement you lost, a decision where reasonable people disagreed, a time you changed your mind, and a time you raised a concern nobody wanted to hear. For each, be ready to say how you felt then and what you'd do differently now.

Hiring manager and project presentation. Choose one project you can explain from start to finish: why it existed, the design, the trade-offs, what went wrong and what you'd change. Practise presenting it, then ask a colleague to question you on any part of it. Be ready, too, to describe how you use AI tools in your own work, including a time you caught a mistake in their output.

What we couldn't confirm

These claims come from a single source each, so we haven't relied on them:

• References and team matching come after the onsite (2026).
• Referred candidates sometimes skip the timed coding assessment (2026).
• Applied AI Engineer and Forward Deployed Engineer interviews may include a customer role-play and a design interview on deploying Claude for a customer (2026).