The short answer: Anthropic does not publish a stage-by-stage research engineer loop. It says interviews are held by video, technical roles use live coding tools, and you can look things up during coding interviews. Several of its research engineer postings say all interviews for the role are in Python. Its candidate guidance says live interviews and take-homes are done without AI unless Anthropic says otherwise, and its engineering blog describes a performance engineering take-home it redesigned three times because Claude kept matching top candidates. Candidate reports describe a recruiter screen, a technical assessment, a hiring manager call and a virtual onsite of about 4 to 5 hours. Last reviewed October 2026.
Anthropic Research Engineer at a Glance
Levels
Research Engineer - Staff / Staff+
Anthropic research engineer interview: quick answers
| Question | Short answer |
|---|---|
| How many interview rounds? | Anthropic does not publish an RE loop. Candidate reports describe a recruiter screen, a technical assessment, a hiring manager call and a virtual onsite of several back-to-back rounds, about 4 to 5 hours in total. |
| What do the rounds test? | Anthropic says technical roles use live coding, and several RE postings say all interviews are in Python. Candidate reports describe ML coding such as attention or training loops, debugging, system design for large-model training, a paper discussion, a project deep dive and a values round. |
| Can I use AI tools? | Only when Anthropic says so. Its guidance says live interviews are "all you" with no AI assistance unless it indicates otherwise, and take-homes are done without Claude unless it indicates otherwise. Its performance engineering take-home explicitly allows AI. |
| How much does an Anthropic research engineer make? | US postings list annual salary ranges from $300K to $400K (Applied AI) up to $500K to $850K for most RL and Staff roles. The postings do not state an equity amount. |
| How long does it take? | Candidate guides disagree: one says about 6 to 10 weeks, another about 20 days. |
| Do I need an ML background? | Not always. Anthropic says about half of its technical staff had no prior ML experience, and its Interpretability RE posting says no research experience is required. |
The Anthropic research engineer role
What Anthropic research engineers work on, the skills its job postings ask for, and what a week can look like. Anthropic posts the role as Research Engineer, <team> (for example Pretraining Scaling, Interpretability, RL Engineering), combined Research Engineer / Research Scientist postings, and Staff or Staff+ Research Engineer.
Primary Focus
What You Need
- Strong software engineering
- Large-scale distributed systems and high-performance computing
- Training, fine-tuning or evaluating large language models
- Debugging training processes
- Python, Rust, Go or Java proficiency (Interpretability)
- 5 to 10+ years building software (Interpretability)
Nice to Have
- PyTorch and CUDA on GPUs, or JAX and XLA on TPUs
- Constitutional AI and RLHF techniques (Post-training)
- Activation extraction and steering infrastructure (Interpretability)
- JAX or PyTorch for RL research (RL Frontiers)
Languages
Tools & Technologies
Want to ace the interview?
Practice with real questions.
A Week in the Life of a Anthropic Research Engineer
Illustrative week · Anthropic
An illustrative week for a research engineer, built from what Anthropic says about the role in its own job postings and publications. Real weeks vary by team.
Culture notes
- Anthropic says engineers do lots of research and researchers do lots of engineering, and all of its papers have engineers as authors.
- Postings describe research as an empirical science done by one team on a few large efforts, with frequent research discussions.
- One company value is to do the simple thing that works, favoring impact over sophistication.
- Some roles carry on-call duty: Pretraining Scaling engineers respond to incidents during launches.
Key takeaway: Expect to do research and engineering in the same week, in Python, on a few large shared efforts.
Anthropic research engineer levels and pay
Your level sets the bar you are interviewed against and the pay band of the offer.
Anthropic Research Engineer Levels
Each level has different expectations, compensation, and interview focus.
What This Level Looks Like
Builds and runs training, evaluation or research infrastructure for one team, such as post-training or interpretability.
Interview Focus at This Level
Candidate reports describe Python ML coding, debugging, large-model system design and a paper discussion.
Find your level
Practice with questions tailored to your target level.
Anthropic does not publish level names. Postings say minimum years of experience track internal job levels, and senior openings carry Staff or Staff+ in the title. The pay table uses posting titles.
Pay in Anthropic's US job postings
Anthropic's US postings list an annual salary range in USD. They mention optional equity donation matching in the benefits text but do not state an equity amount.
| Posting | Location | US salary range |
|---|---|---|
| Applied AI, Research Engineer | San Francisco or New York | $300K to $400K |
| Research Engineer, Interpretability | San Francisco | $315K to $560K |
| Security Engineer / Research Engineer, RL Security | San Francisco, New York or Seattle | $320K to $405K |
| 5 postings: ML/Research Engineer, Safeguards, Research Engineer, Life Sciences, Research Engineer, Production Model Post-Training and others | Several US locations | $350K to $500K |
| 6 postings: Research Engineer, Discovery, Research Engineer, Knowledge Team, Research Engineer, Pretraining Scaling and others | Several US locations | $350K to $850K |
| 12 postings: Research Engineer, Chip Design RL, Research Engineer, Code RL (Reinforcement Learning), Research Engineer, Computer Use and others | Several US locations | $500K to $850K |
Anthropic posts no level names, so each range belongs to one posting. London and Zurich postings show no range. Ranges seen October 29, 2026.
Anthropic research engineer interview process
Anthropic publishes general interview facts on its careers page but no research engineer loop, so the stages below come from candidate reports and guides, with Anthropic's own statements marked.
Screen
3 roundsRecruiter screen
Anthropic says its interviews ask about your experience and what motivates you. Candidate reports say the screen covers your background, motivation and why Anthropic.
Tips for this round
- Have a clear answer to why Anthropic, tied to its research and mission.
- Ask whether your team's technical screen is live, an online assessment or a take-home.
Technical assessment
Anthropic says technical roles use live coding tools, and several RE postings say all interviews are in Python. Candidate reports disagree on the format: some describe a 90-minute online assessment that builds across levels, others a 50 to 60 minute live session, and some teams use take-homes. Reported content includes ML coding such as attention or a training loop, and debugging.
Hiring manager call
One candidate report describes a conversation about past projects, including a review of code samples.
Onsite
1 roundVirtual onsite
Candidate reports describe coding or ML coding with debugging, system design for large-model training, a research or paper discussion, a project deep dive, and a culture and values round.
Tips for this round
- Practice designing large-model training systems: parallelism, checkpointing and failure recovery.
- Read a few recent Anthropic papers closely enough to discuss the claim, the evidence and what you would test next.
- Prepare for a values round on safety and why the mission matters to you.
Tips to Stand Out
- Be fluent in Python syntax and the standard libraries. Anthropic lets you look things up, but says gaps there eat your time.
- Put independent research, thoughtful blog posts or open-source work at the top of your resume, as Anthropic suggests.
- If your background is engineering, apply as an engineer. Anthropic says you will perform better in the interviews.
- Use Claude to research Anthropic and practice answers before interviews, which its guidance encourages, then go without AI in the rounds unless told otherwise.
- Candidate reports say the technical screen format varies by team, so ask your recruiter whether it is a live session, an online assessment or a take-home.
Based on Anthropic's own pages and candidate reports.
Candidate guides put the timeline anywhere from about 20 days to 6 to 10 weeks. Anthropic says it cannot give feedback on resumes or interviews, and that you can re-apply after 12 months. Last reviewed October 2026.
What is different at Anthropic
- Engineers do research here. Anthropic says its engineers do lots of research and its researchers do lots of engineering, and that every Anthropic paper has engineers as authors, often as first author. Many openings are combined Research Engineer / Research Scientist postings.
- Interviews are in Python. Several RE postings say all interviews for the role are conducted in Python. Anthropic says you can look things up during coding interviews, but you should know basic syntax and the standard libraries well enough that they do not slow you down.
- AI is off unless Anthropic says otherwise. Its guidance says live interviews are "all you" and take-homes are done without Claude unless indicated. Its engineering blog says the performance engineering take-home, run since early 2024 and completed by more than 1,000 candidates, explicitly allows AI, and was redesigned three times because Claude models kept matching top candidates.
- Research runs as big science. Postings describe research as an empirical science done by one cohesive team on a few large efforts, with frequent research discussions and a high value on communication. One company value is to do the simple thing that works.
- Some roles carry on-call. Post-training, Pretraining Scaling and Cybersecurity RL postings mention on-call rotation or short-notice incident response, and Pretraining Scaling mentions long days around model launches.
- You need to be in an office at least 25% of the time. Anthropic says most staff are in the Bay Area and come in regularly, and some come in one week a month. US RE postings are in San Francisco, New York and Seattle, and a few are remote-friendly with travel.
Anthropic research engineer interview questions
Questions from our bank, then practice prompts we wrote for this guide around Anthropic's work.
Questions from our bank
From the DataInterview question bank: questions tagged to Anthropic first, then questions that fit research engineer interviews. Open one to answer it.
Multi Head Attention Benefits
Attention Repetition Issue
Offline vs Online Evaluation
Model Artifact Management
Prompt Injection Attack
Differential Privacy Fine Tuning
PPO vs DPO Differences
Tensor Parallelism Comparison
Distillation vs Fine Tuning
ML coding practice
Prompts we wrote for this guide around Anthropic's work. They are not reported interview questions. Each answer is an outline or a worked solution.
In PyTorch, capture the residual stream activations at one layer of a transformer during a forward pass, then steer the model by adding a fixed vector at that layer.
Sample Answer
Register a forward hook on the chosen block that copies its output to a buffer, detached and moved to CPU if memory is tight. For steering, register a hook that returns output + alpha * v, where v is the steering vector, broadcast over batch and positions. Remove hooks in a finally block so they do not leak into later runs. A strong answer checks that alpha = 0 reproduces the original logits exactly and discusses memory when capturing many layers over long sequences.
Implement multi-head self-attention from scratch in NumPy for an input of shape (seq, d_model), and state the shape at each step.
A pretraining run's loss spikes and then turns into NaN a few hundred steps later. Walk through how you would debug it.
Write a function that packs tokenized documents into fixed-length training sequences and builds the mask that stops tokens from attending across document boundaries.
Research discussion practice
Prompts we wrote for this guide around Anthropic's work. They are not reported interview questions. Each answer is an outline or a worked solution.
You add a new principle to a Constitutional AI training setup. How would you check that it changes the intended behavior without making the model less helpful?
Sample Answer
Write targeted evals for the behavior the principle should change and a broad helpfulness eval it should not hurt. Train with and without the principle from the same starting checkpoint and compare both. Look for overcorrection, such as refusing requests that are fine. A strong answer runs the simplest comparison first and reads real samples before trusting the scores.
A steering experiment shows that adding a feature direction changes the model's outputs. How would you check that the effect is real and specific to that feature?
How would you evaluate a model that operates a computer through screenshots and mouse and keyboard actions?
In code RL, the model starts passing tests by special-casing the test inputs. You have two ideas: a complex learned detector, or simple held-out tests. Which do you try first?
Fundamentals practice
Prompts we wrote for this guide around Anthropic's work. They are not reported interview questions. Each answer is an outline or a worked solution.
You have a compute budget of 1e23 FLOPs. Using training compute of 6ND and about 20 tokens per parameter, roughly what model size and token count are compute-optimal?
Sample Answer
Set D = 20N, so C = 6N x 20N = 120N^2. Then N = sqrt(1e23 / 120) = sqrt(8.33e20), about 2.9e10, so roughly 29B parameters. D = 20N, about 5.8e11, or 580B tokens. Check: 6 x 2.9e10 x 5.8e11 is about 1.0e23.
When would you use data parallelism, tensor parallelism and pipeline parallelism?
A new tokenizer averages 4.0 characters per token instead of 3.2. What does that change for training and serving?
Why do training runs save checkpoints often, and what goes into one?
Behavioral practice
Prompts we wrote for this guide around Anthropic's work. They are not reported interview questions. Each answer is an outline or a worked solution.
Why do you want to work on AI safety at Anthropic?
Sample Answer
Give a specific reason tied to Anthropic's research, such as a paper or a team's work, and what you would contribute. Show that you have thought about the risks and not only the technology.
Tell me about a time you chose a simple solution over a sophisticated one.
Tell me about a result that turned out to be wrong.
Describe a time you explained a technical result to people outside your team.
Practice problem: find the rollback point after a loss spike
pythonA pretraining job logs (step, loss) pairs in step order and saves checkpoints at known steps. Write find_rollback(losses, checkpoints, window, factor). A step is a spike if its loss is NaN or infinite, or if at least window earlier losses exist and its loss is greater than factor times the mean of the previous window losses. Return (spike_step, rollback_step) for the first spike, where rollback_step is the latest checkpoint strictly before the spike, or None if there is none. Return None if no step spikes. Example: losses [(1, 2.0), (2, 1.9), (3, 1.8), (4, 1.7), (5, 4.0)], checkpoints [2, 4], window 3, factor 1.5 returns (5, 4), since 4.0 is more than 1.5 times the mean 1.8.
1from typing import List, Optional, Tuple
2
3def find_rollback(losses: List[Tuple[int, float]], checkpoints: List[int], window: int,
4 factor: float) -> Optional[Tuple[int, Optional[int]]]:
5 """Return (spike_step, rollback_step) for the first loss spike, or None."""
6 pass700+ coding problems with a live code executor.
Practice in the EngineHow to prepare
A plan for the Anthropic loop, most important first:
- Do every practice problem in Python, since several Anthropic RE postings say all interviews are in Python, and get fluent enough with the standard library that look-ups are rare.
- Practice ML coding from scratch: attention, a training loop, data packing with masks, and activation hooks in PyTorch.
- Practice coding problems that build in stages, getting a working version first and then extending it, without AI tools.
- Read several recent Anthropic papers and be ready to discuss the claim, the evidence and the next experiment you would run.
- Prepare a system design for large-model training: parallelism, checkpointing, failure recovery and what breaks at scale.
- Prepare stories for a values round: why safety matters to you, a simple solution you chose, and a result you got wrong.
Practice path
Courses that cover the rounds above:
- Anthropic Research Engineer Interview
- LLM Architecture
- PyTorch
- Deep Learning
- GPU Programming
- ML System Design
Are you ready for the Anthropic research engineer interview?
1 / 6Can you implement attention, a training loop or a tokenizer from scratch in PyTorch or NumPy, and debug it?
Frequently Asked Questions
How many rounds is the Anthropic research engineer interview?
Anthropic does not publish the loop. Candidate reports describe a recruiter screen of about 30 minutes, a Python technical assessment, a hiring manager call of about an hour, and a virtual onsite of about 4 to 5 hours with several back-to-back rounds, then references.
Can I use AI tools in an Anthropic research engineer interview?
Only if Anthropic tells you so. Its candidate guidance says live interviews are "all you" with no AI assistance unless it indicates otherwise, and that take-homes are completed without Claude unless it indicates otherwise. Its performance engineering take-home explicitly allows AI. For applications, Anthropic asks you to write the first draft yourself and then use Claude to refine it, and it suggests using Claude to prepare.
What language are Anthropic research engineer interviews in?
Python. Several RE postings say all interviews for the role are conducted in Python. Anthropic says you may look things up during coding interviews, but you should know the syntax and standard libraries well.
How much do Anthropic research engineers make?
In US postings seen in October 2026, most RL, evaluation and Staff RE roles list $500K to $850K. Pretraining, Discovery and Knowledge roles list $350K to $850K, Alignment and Post-Training list $350K to $500K, Interpretability lists $315K to $560K, and Applied AI lists $300K to $400K. London and Zurich postings show no range.
Do I need a PhD or ML experience to be an Anthropic research engineer?
No. Anthropic says about half of its technical staff have PhDs and about half had no prior ML experience. Its Interpretability RE posting asks for 5 to 10+ years building software and says no research experience is required. If you have an engineering background, Anthropic suggests applying as an engineer.
Does Anthropic offer internships or research programs?
Anthropic says it does not offer internships. The Anthropic Fellows Program is 4 months of full-time research with mentorship, a weekly stipend of $3,850 and about $15K a month of compute, with the next cohort expected in January 2027. The Research Residency is a paid six-week to six-month program for professors, technical founders or staff-level engineers moving into AI, with applications closing November 22, 2026.
Is the Anthropic research engineer role remote?
Mostly not. Anthropic expects staff in an office at least 25% of the time. A few RE postings are listed as remote-friendly with travel required, and some roles include on-call or short-notice incident response.
