Anthropic Phone Screen Interview: What Gets Tested and How to Pass

Anthropic runs one of the stranger phone screen interview processes in tech. Before you meet a live interviewer, you clear two distinct gates: a recruiter call that is explicitly not a formality, and a CodeSignal assessment built around one escalating problem. Get both right and you're talking to the hiring manager. Fumble either and the loop ends. No second chances, no "we'll keep your resume on file."

This guide covers both stages: what's actually evaluated, what the problems look like, how scoring works, and what separates candidates who advance from candidates who are currently refreshing their inbox.

The Recruiter Call Isn't a Formality (No, Really)

Most companies treat the recruiter call as scheduling overhead. You discuss your background, they explain the role, you both pretend to care about each other's timezone preferences. Anthropic doesn't do that.

The call runs 30 minutes. Some candidates get it before the CodeSignal link. Others receive the assessment immediately after applying and take the call afterward. The order varies by role. Either way, the recruiter screen is a pass/fail filter, not a warm-up.

A former Anthropic recruiter told candidates directly that "so many people get rejected because of the behavioral aspect, even if they've done really well technically." That pattern starts here, not at the culture round weeks later. The recruiter is not a gatekeeper you charm. They are an interviewer.

Three things the recruiter is listening for:

Your actual understanding of what Anthropic does. Generic answers about "working on AI" don't land. Reference Constitutional AI, interpretability research, or frontier model safety. You don't need a PhD-level position. You need to have thought about it, out loud, before this call.

Why this company specifically. "I want to work on interesting problems" describes every AI company, every startup, and your cousin's NFT venture. Anthropic expects candidates to have a view on the distinction between building capable AI and building safe AI. Come with a formed opinion.

Engineering signal. The recruiter will ask about past work. Pick one project where your individual impact is clear, not a team contribution where you'd need a paragraph to explain your slice. "We shipped a distributed cache and I owned the eviction policy" beats "we built a microservices platform" every time.

There's no coding on this call. Going in without a real position on the mission question is the most common early exit.

835 Out of 850. That's the Bar.

After the recruiter call, you get a link to a 90-minute asynchronous CodeSignal assessment. This uses the General Coding Framework (GCF): one complex problem unfolding across four progressive levels. Scoring runs 600 to 850, based on correctness, code quality, and time efficiency.

Community reports from 2025 and 2026 Anthropic candidates put the cutoff for software engineering roles at approximately 835 out of 850. That leaves 15 points of margin for error. If you've ever taken a standardized test and thought "I need to do well," forget that. This is "I need to do nearly perfectly."

The upshot is that the format rewards a very specific skill: building a clean, working system one layer at a time under time pressure. Not algorithmic wizardry. Not trivia. System construction.

Anthropic's problems are not LeetCode-style. You won't see "find the k-th largest element." The problems are practical and stateful, designed to look like building a small system one feature at a time. If you have done nothing but grind two-sum variants, you will find level 3 of this assessment deeply unpleasant.

Problem shapes that show up repeatedly:

In-memory database. Level 1: implement SET and GET. Level 2: add TTL. Level 3: transactions (BEGIN, COMMIT, ROLLBACK). Level 4: prefix scanning or filtering. Basically: build a lite version of Redis, in Python, in 90 minutes.

Web crawler. Level 1: synchronous BFS from a seed URL. Level 2: extract links, handle revisit cycles. Level 3: multi-threaded crawling with a thread pool. Level 4: reason about the GIL or redesign with asyncio.

Bank or transaction system. Level 1: deposits and withdrawals. Level 2: transfers. Level 3: rollback on failure. Level 4: auditing or concurrent access.

Each level builds on the previous one. If you stall on level 2, you don't see level 3. Scoring is heavily weighted toward reaching later levels cleanly. A mediocre level 4 submission beats a polished level 2 submission.

How Your Score Gets Calculated

Three things feed into it:

Correctness. Tests pass or they don't.
Code quality. The GCF system penalizes messy code even when output is correct. One giant function that passes tests loses points here. Name your variables like a person, not like a person who is running out of time.
Time efficiency. Moving forward quickly rewards your score. Running out the clock on level 2 hurts it.

Python is the right choice. Anthropic's engineering culture is Python-first, and community reports suggest Python submissions score better on code quality criteria. More practically, the standard library covers most of what you'll need: collections.defaultdict, heapq, threading.Lock, asyncio.Queue. Using them correctly is itself a signal.

How to Not Implode Under Pressure

Start minimal, not complete. Implement the simplest version that passes level 1 before adding anything. Don't architect the full system upfront. You don't know what level 4 will ask, and you'll spend 30 minutes designing abstractions for a problem that doesn't exist.

Test as you go. The GCF environment supports running tests mid-session. Candidates who test incrementally catch bugs before they compound into a level 2 failure.

Write clean code, not clever code. One obvious, well-named solution beats a compact one that requires a comment to understand. The GCF grader is not impressed by your one-liner.

Know Python concurrency basics. Levels 3 and 4 often surface threading or async. You don't need to implement a production thread pool. But you need to reason correctly: threading works for I/O-bound work despite the GIL (it releases on blocking I/O), multiprocessing for CPU-bound work, asyncio for high-concurrency I/O without thread overhead.

The Hiring Manager Call: No Coding, Plenty of Scrutiny

Clear the CodeSignal and you'll land a hiring manager call: 45 to 60 minutes, focused on engineering judgment, project depth, and technical decision-making. No coding.

Pick a project where you made the architecture calls. Not "the team built X." The hiring manager wants your specific decisions: the options you weighed, why you picked what you picked, what broke, what you'd change. If you can't trace the reasoning behind the choices, pick a different project. If you were a junior who shipped features from a spec, pick a different project.

The AI safety thread from the recruiter call continues here. Expect questions about how you think about failure modes in systems you've built. Not philosophical. Practical: how did you test it, what were the edge cases with real-world consequences, how did you reason about correctness under uncertainty.

Where the Phone Screen Fits in the Full Loop

The two screening stages sit inside a larger loop. Knowing the full shape helps.

Stage	Duration	Format
Recruiter screen	30 min	Call
CodeSignal assessment	90 min	Async, self-scheduled
Hiring manager screen	45-60 min	Call
Onsite loop 1	2-3 rounds	System design, coding, culture
Onsite loop 2	2-3 rounds	Experience, behavioral, values

Onsite loop 2 only happens if you pass loop 1. Anthropic makes an explicit pass/fail decision after the first onsite day. The full process typically takes four to six weeks from first contact to offer decision.

For a deeper look at what the onsite rounds test, see the Anthropic software engineer interview guide.

What Actually Knocks People Out

Showing up to the recruiter call without a view on AI safety. You don't need a published position. "I think responsible development matters, and here's why Anthropic's approach is interesting to me" is fine. Generic enthusiasm is not. "I'm really excited about AI" will get you out of this call faster than a wrong answer.

Treating the CodeSignal assessment like a speed contest. The GCF format rewards incremental, clean progress over burst cleverness. A candidate who solves level 3 with readable code scores higher than one who reaches level 4 with code nobody would maintain. This is not LeetCode. Slow down and write things properly.

Choosing the wrong project for the hiring manager call. If your strongest work was a team effort where your individual decisions are hard to reconstruct, you'll struggle with follow-ups. Pick something you can own end to end. "We" is a red flag. "I" is what they're listening for.

Conflating "excited about AI" with mission alignment. Anthropic is not just an AI company. It's specifically an AI safety company, and that distinction matters to everyone you'll meet there. See how the interviews differ at comparable labs in the OpenAI vs Anthropic interview comparison.

How to Prepare for Each Stage

Recruiter screen: Read Anthropic's Constitutional AI overview and skim the interpretability research page. Come with a formed view on why safety and capability are sometimes in tension. You're not expected to resolve this. You're expected to have thought about it, and to be able to talk about it for three minutes without retreating to platitudes.

CodeSignal: Practice building small stateful systems in Python, one feature at a time. LRU cache, LFU cache, task scheduler with dependencies, in-memory key-value stores: these are closer to what you'll see than standard algorithm problems. Do timed sessions where you add features progressively. See Python for coding interviews for a standard library reference.

Specific things to drill: threading.Lock for shared mutable state, collections.OrderedDict for insertion-order maps, BFS with deque, and TTL expiry logic using timestamps.

Hiring manager call: Write down two or three projects where you made architecture decisions. For each: what were the options, why did you choose what you chose, what broke, what would you change. Keep answers concrete. "We chose Postgres over Redis because writes were infrequent and we needed transactions" beats "we evaluated several options and picked what made sense."

The gap between understanding a decision and articulating it under pressure is real, and repetition is the only way to close it. SpaceComplexity runs voice-based mock interviews with rubric feedback, which simulates the Anthropic hiring manager call far better than solving problems in a text editor.

Before any live call, work through asking good clarifying questions. The habit pays in both the technical and hiring manager stages.