The Anthropic Internship & New-Grad Interview

The Anthropic new-grad and internship software engineer interview runs on the same spine as the full-time loop. Expect a practical CodeSignal build the recruiter calls "practical, not LeetCode," a dedicated Values round that reads how you think about AI safety, and a strict no-AI rule in the room. There is no separate algorithm-heavy early-career track. If you prepared by grinding three hundred LeetCode problems, you prepared for a different company's interview.

One honest caveat before anything else. Anthropic publishes very little about a dedicated internship or new-grad pipeline, and early-career headcount is limited compared with the frontier labs that run large university programs. Where this guide is confident, it is because the coding format, the no-AI policy, and the seven values are stated by Anthropic or consistent across candidate accounts. Where a claim would be intern-specific and unverified, this lesson flags it rather than inventing a number. Treat what follows as the established core of the loop applied to an early-career candidate, not a leaked intern rubric.

The Anthropic angle

The AI lab that builds Claude bars Claude, and every other assistant, from its own interviews. Anthropic's candidate guidance is explicit: live screens and take-homes are you alone, no AI help, unless an interviewer says otherwise. For a new grad used to an editor with autocomplete and a chat window open, that is the single biggest adjustment. Anthropic also reportedly screens for code engineered to slip past the visible tests rather than genuinely solve the problem. So practice the way you will be assessed, from a blank file and your own memory, and the rule feels normal on the day instead of like a missing limb.

The practical build screen does the filtering

For an early-career candidate the technical screen is the gate that matters most. It is a CodeSignal build. One problem grows across four locked levels in about ninety minutes. Level one might be SET, GET, and DELETE on an in-memory key-value store. Later levels bolt prefix scans, time-to-live expiry, then file compaction and a capacity limit onto that same store. The methods you wrote at level one still have to keep passing through all of it. The reported bar is roughly 520 of 600 points, with at least three of the four levels cleared, and each level unlocks only after every hidden test for the one before it turns green. Running out of clock before the last level is the common miss.

Because the live bank is small and repeats, breadth is not the test. The full mechanics of beating the grader, level by level, live in The Anthropic Online Assessment. The recurring problems, the key-value store, the banking ledger, a Premium question, are worked in The Anthropic Coding Interview. What gets graded is clean, modular code. Beyond correctness, it is the edge cases you catch on your own initiative and the tests you write before anyone asks. It is also whether your design absorbs the next constraint without a rewrite when the interviewer stacks one on. A passing first solution is only the starting line.

What the loop tests, and what carries the most weight

Pass the screen and you reach a small onsite block, reported at roughly four to five hours. Expect live coding in shared Python with escalating follow-ups, often a system design discussion, a project deep-dive, and a standalone Values round. The exact mix flexes by team and by how much scope you can show. The design round leans lighter for a candidate who has not architected large systems yet. What does not get lighter is the Values round. It is scored the same way for an early-career candidate as for a senior one.

• The gate: CodeSignal build. One problem, four locked levels, about 90 minutes, near 520 of 600 to advance and at least three levels cleared. Practical and stateful, not an algorithm quiz. Some pipelines run a 60-minute live screen instead, most often a Premium question.
• The real filter: The Values round. A dedicated round with non-technical interviewers, reported as the most common reason technically strong engineers get cut. It reads how you reason about AI risk and benefit under pressure, and it is weighted the same for a new grad as for a senior.
• The rule: No AI assistance. Live rounds and take-homes are unaided unless an interviewer says otherwise, and gaming the tests reportedly gets flagged. Prep with Claude all you want; in the room it is you and a blank file.
• Language: Python expected. Python is the default and the standard library is rewarded. Other languages have been allowed, so plan around Python unless your recruiter says otherwise.

The Anthropic angle

The highest-leverage thing an early-career candidate can do is take the Values round as seriously as the code. A student's instinct is to pour every hour into coding and treat the culture conversation as a formality, which is exactly backwards here. Weight it like a technical round. The round grades whether you can hold AI's real risk and its real benefit at once, what Anthropic calls "hold light and shade," and it rewards honest skepticism over rehearsed enthusiasm. Come with one genuine view on AI safety you reasoned to yourself, anchored in something you read. The full preparation is in The Anthropic Culture Interview & Values Round.

Where a new grad lands, and how the offer is built

Anthropic uses plain titles, Software Engineer, Senior, Staff, not a numbered ladder. Level is read off the scope you demonstrate. A new grad enters at the early Software Engineer rung. Learn one mechanism early: team matching. Even a clean loop is not an offer until a team has headcount for you, which is the main reason a strong candidacy can go quiet for weeks. That silence is not a rejection. On the money side, an offer is base plus double-trigger private RSUs, restricted stock in a company that has not gone public, with little to no cash bonus. How that equity actually works is in the negotiation lesson.

Common questions