Interview in 2-4 weeks

With 2-4 weeks left, don’t work through the whole library. Pick your company’s page below, do the 10-15 Sandbox problems it names, practise the way that company runs its interview (code run or not, timed levels, two problems in 40 minutes), and give the behavioral round real hours. For most companies it decides more than people expect.

The companies side by side

	Coding style	Is your code run?	AI in interviews	Behavioral weight
Anthropic	Practical, multi-level builds (3-4 levels), mostly Python	Yes (CodeSignal / Colab)	No, unless told; experiments in 2026	High: the values round is where most fail
OpenAI	Practical multi-part problems with tests; algorithms inside	Yes (CoderPad)	“Varies by interview”	Medium-high: deep-dive and “why OpenAI”
Google	One medium-hard algorithm problem per 45 min, optimal, DP in scope	No: trace by hand	No	Medium: can sink you
Meta	Two problems in 40 min, broad, no DP; plus an AI-enabled multi-file round	No (except the AI round)	Banned, except the AI-enabled round	High: failing it is a no-hire
Amazon	Medium problems, class-based / low-level design, no pseudocode	No	No (press-reported)	Highest: LPs in every round
Stripe	Practical coding, bug bash in a real repo, API integration	Yes, on your own machine	No, except a 2026 AI pilot round	Medium-high: operating principles
Trading firms	One problem that grows; process over finish; systems depth at HRT and Citadel	Usually (CoderPad)	No published policy	Low: a project deep-dive instead

Each page tags every claim as official (the company’s own words), reported (two or more independent candidate reports) or thin (one source).

What every company tests

These came up in the evidence for every company. If you only have a few days, do these:

1. Clarify before coding. Ask about inputs, scale, edge cases and what to return. Every official guide asks for it. See Edge cases.
2. Think aloud. Silence is the most-cited failure reason everywhere.
3. Get a working baseline, then improve it. At Anthropic, OpenAI and Stripe the problem grows, so a clean Level 1 beats a clever half-solution. See Incremental requirements.
4. Verify your own code. Trace an example, then test edge cases, before the interviewer does it for you. Meta and Amazon score this explicitly. See Testing and Debugging.
5. State the complexity, and know what you’d change at 10x the data. See Complexity analysis.
6. Have project stories with numbers: what you decided, the alternatives, what you’d change.
7. Practise without AI and without autocomplete, unless the round says otherwise. And ask your recruiter, round by round, because policies changed during 2025-26.

How to use the last 2-4 weeks

1. Day 1: find out the format. Ask your recruiter: which rounds, which language, is code run, is AI allowed anywhere, in person or remote. Your company page lists the exact questions.
2. Match the conditions. Practise in the Sandbox under your company’s rules. If its editor doesn’t run code (Google, Meta, Amazon), trace by hand before you press Run and count your runs. If it runs timed levels (Anthropic, OpenAI), set a timer and keep every level passing.
3. Do the short list, not the library. Each company page names 10-15 problems mapped to what candidates actually report. When a problem exposes a gap, read that one Study page and redo the problem from scratch two days later.
4. Split the time. For roughly 4 weeks: 60% coding, 20% design (if your level has it), 20% behavioral. For Amazon, make behavioral 35%.
5. Mock the real thing in the last week. Do two full-length timed mocks, one behavioral mock with a friend, and nothing new the day before.

The evidence has limits

No company publishes how often each topic comes up. Candidate reports are biased toward people who post online, and formats vary by team, level and country. These pages say where the evidence is thin or conflicting. Your recruiter’s description of your loop always beats this page.

Pick your company

AnthropicMulti-level builds, concurrency, values round

OpenAIPractical multi-part coding, deep-dive

GoogleOptimal algorithms, no execution, committee

MetaTwo problems in 40 min, AI-enabled round

AmazonLeadership Principles, robust code, low-level design

StripeBug bash, integration, idempotency

Trading firmsJane Street, Citadel Securities, HRT