Anthropic Software Engineer Interview Experience — A Compilation on Culture, Practical Coding, and Timelines

A clarification: this is a compilation, not my own interview experience. The content comes from public interview reports on this forum, Blind, Glassdoor, and Interviewing. It took me quite a long time to organize. Anyone who's interviewed there is welcome to add information or corrections. There's a forum compilation of 25 posts with a very complete question list, so I won't repeat the questions here. Here's the reference:
25-post compilation
The easiest place to trip up at this company this year is the culture round. It rejects more people than coding does, so I'll start there.
Several posts say this round eliminates the most people, and recruiters say the same thing. Its style is strange. In Interviewing's article interviewing Anthropic engineers, a candidate described the round as therapy: they ask how you felt at the time, what you thought afterward, and keep digging. The STAR formula is basically no use here.
There was a Senior Staff infrastructure applicant on Blind whose coding question involved fixing an instruction loop in a bootloader, whose system design was a sharded model downloader, and whose deep dive kept pressing whether a project that consumed 50 engineer-years was worth 20 million. They said the technical rounds went all right and ultimately suspected this round was where they failed. Questions included things like, "What kind of work do you hate most? If someone else hates it too, how do you persuade them to do it?" If you haven't thought about it, you really can't make up an answer on the spot.
A few questions match across the forum and Glassdoor: Have you faced an ethical conflict at work, and how did you handle it? Where do you disagree with Anthropic's direction (Glassdoor's exact wording: What ways do you disagree with our AI approach)? If shipping speed conflicts with safety, which do you choose?
From what I've read, agreeing with everything is actually dangerous. They clearly want to see whether you have your own judgment. If every answer is "You're right," this round is a lost cause. There's also a claim that when an interviewer challenges you, being persuaded by a good argument and changing your answer on the spot doesn't seem to cost points. I couldn't find a second source for that, so I'm leaving it here with a question mark.
There's another hurdle earlier on: the recruiter screen.
It's only 30 minutes, but the filtering is harsh. The forum compilation included someone with 10 years of experience whose answers on "why Anthropic" and safety were too shallow, and they were out at this stage. The questionnaire asks roughly: How is Anthropic different from other labs? What happens if AI is misused? What are advanced AI's biggest risks and benefits? A generic answer like "I want to work on cutting-edge AI" might work elsewhere, but not here.
Something else many people don't know: the online application has a written "why Anthropic" question, 200–400 words, and the company explicitly says it matters a lot. I'll also note the AI rules to help avoid missteps: you can write application materials yourself first and then use Claude to polish them, and you can use it to prepare; take-homes and live interviews prohibit it by default unless that round's instructions explicitly allow it.
Coding isn't tested the LeetCode way.
The OA is on CodeSignal: four consecutive levels building a system, such as an in-memory database or banking system. Passing one level unlocks the next. Plenty of people run out of time rather than not knowing how to solve it. The onsite is more practical: an image pipeline uses PIL for resizing and rotation, then adds multiprocessing; a crawler starts single-threaded and becomes concurrent, and thread safety for visited is almost always asked. Hack2hire went through 15 interview reports and said these questions are designed around using libraries, with little room for pure data structures. Sometimes no test cases are provided, so you need to create and run your own. Explain as you write; quietly finishing everything before submitting can actually score poorly.
The question pool really is small. Someone on LeetCode Discuss said the phone-screen and system-design pools were absurdly small. Their phone-screen stack-trace question was the exact type they'd practiced beforehand. Original post:
LeetCode Discuss original post
Another counterintuitive thing about system design: it isn't drawing on a whiteboard; it's working through the design in writing in a Google Doc. The Staff post on Blind had Collaborative Prompt Playground, with collaboration, versioning, and execution flows all expressed in text. If you've practiced the usual big-tech boxes-and-arrows approach, you're heading in the wrong direction.
There was also a new agentic-coding round in 2026, described in a separate forum post: a GitHub-like site, four PRs, and Claude Code to make changes, with no switching languages. The author failed.
The original post is here.
There are too few public details about this round, so I won't go further.
Two timeline accounts don't agree, so I'm including both: the forum compilation says gaps of 2–3 weeks between rounds and 1–2 months overall are normal; Interviewing says 3–4 weeks overall, with acceleration possible if you have a competing offer. They agree that rejection emails don't provide feedback and that team matching follows a passed onsite. A Blind post said the manager mentioned beforehand that there were very few openings left, and it still didn't work out in the end.
That's about it. If you've interviewed there, add information or corrections in the comments.