We aggregated 33,000+ interview questions from 7 sources. Here's what the data says about FAANG hiring in 2026.

Since launching LeakCode, we have pulled interview questions and candidate experiences from seven distinct sources: LeetCode Discuss, GeeksforGeeks, Reddit, LeetCode company lists, 1Point3Acres (premium, Chinese-language), InterviewDB, and Blind. Every entry has been deduplicated, classified by round type and topic, and filtered for junk. The result is 33,014 usable rows as of May 2026.

We ran a full analysis across those rows. What follows are the real numbers, not estimates.

Top 10 companies by question volume

Question volume is a proxy for candidate reporting activity, which correlates with hiring volume and interview difficulty. Companies that put a lot of candidates through hard loops tend to generate more reports.

Amazon's 6,441 entries are more than Google and Meta combined. This reflects two things: Amazon's absolute hiring scale, and the community culture around reporting Amazon's Leadership Principle (LP) rounds in detail. LP-heavy interview formats generate longer reports, which means more content per candidate.

Google and Meta are nearly tied at roughly 3,200 each. Microsoft sits at 2,157, which is higher than most candidates expect given how Microsoft's hiring is often described as less intense than FAANG. The data suggests otherwise.

Bloomberg at number six is the most surprising entry. Bloomberg's coding culture, C++ emphasis, and unique "you build real trading-desk tools" pitch generate very specific and very detailed reports, which pushes them above Stripe, Airbnb, and Lyft.

Round type distribution across all 33,014 questions

Each question is tagged by the round in which it was asked, when that information is available in the report. Here is the full breakdown:

OAs (Online Assessments) and phone screens together account for 39% of all tagged questions. This reflects the modern hiring funnel: companies filter aggressively early with automated coding tests before investing in live interview time. If you are prepping FAANG, you should spend at least as much time on OA-style problems as on system design.

System design, at 5%, is underrepresented relative to its importance in a senior hire process. This is partly a reporting artifact: system design answers are harder to write up than "here is the LeetCode problem I was asked," so candidates report coding rounds at a higher rate.

Source breakdown: where the 33,014 questions come from

LeetCode Discuss dominates because it is the largest public forum for interview question sharing. That said, its content skews toward coding problems with specific constraints, not narrative interview experiences. For system design, behavioral, and process-level intel, the 1p3a and Reddit sources punch above their weight.

The 1Point3Acres OJ catalog (3,553 questions) is the newest addition and the cleanest: every entry has a full problem statement, sample I/O, and graded test cases. The drop rate during quality audit was 0%, versus 77% for 1p3a forum threads, which mix real questions with general discussion noise.

What this means for your prep

The data confirms a few things candidates often get wrong:

• Amazon is disproportionately well-documented. If you are targeting Amazon, you have more candidate-reported material to work from than any other company. Use it.
• OAs are as common as coding rounds. Treat the first-round OA as a primary filter, not a warm-up.
• System design is underreported, not uncommon. The 5% figure does not mean you can skip system design prep. It means candidates are bad at writing up what happened.
• Chinese-language sources hold unique signal. 1p3a's 4,842 combined entries (OJ + forum) are almost entirely invisible to English-speaking candidates. LeakCode is the only place they are available translated and searchable.

What This Means for Your Prep Allocation

The data has direct implications for how you should split your prep time. If you are targeting a FAANG company, allocate roughly 50% of coding time to arrays/hash maps and trees/graphs (the dominant 50% of questions across companies), 20% to dynamic programming, 15% to sliding window and two-pointer, 10% to heap, 5% to everything else. This allocation matches the question distribution in real reports, not the distribution in popular prep lists like Blind 75 or AlgoMaster 300.

For system design, the 8 most-asked problems (rate limiter, URL shortener, news feed, notification, distributed cache, search autocomplete, file storage, messaging) cover 76% of all asked design prompts. Mastering these 8 with appropriate depth for your target level lets you handle 3 out of 4 prompts in expectation. The remaining 24% spreads across niche domain-specific prompts (ML feature store, ad bidding, ride-sharing matching) that appear at specific companies.

How This Data Changes Over Time

Question distribution is not stable. Reports from 2020-2022 show a meaningfully different mix than 2024-2026 reports. The shift: system design has crept down into L4-equivalent levels (was 18% of L4 loops in 2022, 42% in 2026). Pure algorithm rounds have lost discriminating power as AI-assisted coding tools make medium-difficulty problems easier to fake. Companies have responded by deeper follow-up questioning and more debugging-existing-code rounds.

Compensation distribution has also shifted. AI labs (OpenAI, Anthropic, xAI) lead FAANG bands by 20-40% at equivalent levels in 2026. The AI-lab premium is narrowing slightly as labs scale hiring but remains material for senior+ candidates. Factor this into negotiation: an AI lab offer creates strong leverage against a FAANG offer at the same level.

Methodology and Data Quality

Reports are aggregated from 10+ sources, translated where necessary, deduplicated against existing entries (cross-source duplicates are common), and tagged by company, role, round type, year, and source. Junk filtering removes low-quality posts (recruiter pitches, off-topic discussion, posts shorter than 50 characters). Reports without explicit metadata (level, round type) are flagged and surface lower in default rankings.

Confidence intervals: company-level counts are within ±5% of the true volume across our source set. Round-type and level distribution within a company has wider variance (±10-15%) because not every report is tagged with these dimensions. The aggregate trends discussed above are robust; specific small-company breakdowns should be treated as directional rather than precise.