Anthropic · Software engineer / research engineer · All levels
Onsite live coding
Live over Google Meet on Colab or CodeSignal with an interviewer watching you think
What this round is judged on
- Production-quality code · readable, correct, sensibly structured
- Debugging speed · locating and fixing defects live
- Design decisions · reasonable tradeoffs on state, API and persistence
- Communication · narrating your reasoning as you build
Where this round sits
Anthropic’s published process. Lime marks the stage this round runs.
- 01Recruiter screen
- 02Technical screen
- 03Onsite loopthis round
- 04Decision and team match
What happens in each stage›
- Recruiter screen
- A spoken call on your background and what motivates you. Anthropic says its interviews cover your experience and your motivations, and that time is set aside for your questions. It does not publish a duration.
- Technical screen
- A live technical interview for most roles, run over Google Meet on Colab or CodeSignal, with reference materials allowed. Some teams use a take-home instead: Anthropic has published one that runs two hours and, unusually, expressly allows AI tools.
- Onsite loop
- Further rounds on coding, on work you have done, and on motivation. Anthropic does not publish how many rounds there are, in what order, or over how many days, and candidate accounts disagree, so treat the shape of this stage as unsettled.
- Decision and team match
- Final decision and matching to a team. Anthropic does not publish how this step works.
Sources
Built from what Anthropic publishes about its own process and from real interview data. No leaked question lists. The questions you get are generated against your own resume, so they are not from anyone else's interview.
Seven guiding principles used as the values rubric; capability over credentials ("We care about what you can do, not where you learned to do it"); engineers do research and researchers do engineering; all interviews over Google Meet; live coding on Colab and CodeSignal with reference materials allowed and basic syntax expected; interviews are explicitly two-way
The current candidate AI policy, last updated 2025-07-10: Claude encouraged to refine a draft you wrote yourself and to prepare for interviews ("research Anthropic, practice your answers, and prepare questions for us"), not allowed in live interviews, and not allowed on take-homes unless Anthropic indicates otherwise
A real Anthropic take-home (2 hours, originally 4, over 1,000 candidates since early 2024); "a normal 50 minute interview" as the standard live slot; live questions built from real but niche work problems, kept in rotation for years; the take-home no-AI default and the explicit per-assessment exception
Independent 2026 reporting that the take-home is current and repeatedly redesigned as Claude improves, and that AI use on that assessment is expressly permitted
That the policy changed: a blanket ban on candidate AI use in early 2025 was replaced by the current staged guidance in July 2025, with Anthropic's head of talent on record
The Anthropic engineering ladder the level field maps to: Software Engineer, Senior, Lead, Staff
- companyAnthropic job posting (Research Engineer, Production Model Post-Training)job-boards.greenhouse.io
"For this role, we conduct all interviews in Python" on research and research-engineering roles; the application-stage AI policy gate linking to the candidate AI guidelines; the weighted "Why do you want to work at Anthropic?" question, present on 43 of 49 live postings
Sources last checked . Hiring loops change, so this date is the honest limit on everything above.
Other Anthropic rounds
- Recruiter screen
Conversational, low-pressure, two-way
- CodeSignal technical screen
Live coding over Google Meet on Colab or CodeSignal, with reference materials allowed
- Technical project deep-dive
You present work you owned end to end while an interviewer probes the decisions
- Values and mission
A conversation about the mission, judgment, and how you work with people