LeetCode isn’t dead, but it’s not the whole picture anymore

Part one

What’s changing

The algorithm round is still standing

interviewing.io asked its interviewers how AI was changing hiring at their companies and got 67 responses, most from FAANG and FAANG-adjacent companies. Of the 52 at FAANG, none said their company had moved away from algorithmic questions. techinterview.org puts it in a sentence: “The bar has risen, but the format has not been replaced.”

Meta is the clearest example. Its new AI-enabled round replaces one of two onsite coding interviews. The other, according to Hello Interview’s guide, is still “One classic LeetCode-style algorithm problem (no AI).” CoderPad’s 2026 hiring survey found the same thing across the industry: “algorithm-heavy tests remain widespread.”

Look at the second number. The interviewers who kept the algorithm round changed what happens inside it. Startups went further: two-thirds of startup respondents said AI had meaningfully changed their process, against none at FAANG and FAANG-adjacent companies. Exponent’s Jacob Simon told LeadDev that companies are still asking LeetCode questions, then added the part that matters: “They’re maybe not weighing them as heavily as they were before.”

Why the old question stopped working

interviewing.io ran the experiment. Candidates leaned on ChatGPT during mock interviews, and the interviewers weren’t told. On verbatim LeetCode questions 73% passed, well above the platform’s usual 53%. Lightly modified questions barely helped. Only questions written from scratch held up, and no interviewer reported suspecting anything. The authors’ conclusion: “ChatGPT has made verbatim questions obsolete.”

Canva reached the same conclusion from the inside. “AI assistants can trivially solve traditional coding interview questions,” its engineering team wrote, before rebuilding its interview around that fact.

What grew up around it

The response wasn’t one new format. It was five or six, and a 2026 loop usually combines several of them.

Multi-file projects

CoderPad’s guidance to interviewers starts here: “Always use a multi-file project template instead of a single-file pad.” Its reasoning is blunt: “Large language models perform significantly worse when they must reason across multiple interdependent files.” That is a vendor’s claim with no data attached, but the industry is acting on it. Meta’s round hands candidates a project that is “multi-file, with existing classes, data models, and logic already written.” Karat’s NextGen interviews use “complex, multi-file projects with an integrated AI assistant.” HackerRank says it in one line: “A code repository is the foundation of your interviews.”

Reasoning over recall

The second instruction in CoderPad’s playbook: “Structure interviews to surface thinking, trade-offs, and adaptability rather than polished final answers.” Interviewers were already there. One at Meta told interviewing.io they are now “more focused on the WHY than the HOW.” HackerRank’s AI interviewer is built to probe the same way, “asking follow-up questions to see how candidates think.”

Requirements that move

A finished answer is no longer the end of the question. CoderPad’s best practices include “Multi-part or progressive problems that evolve over time” and “Introducing follow-up changes mid-solution to test flexibility.” Meta’s round runs in three phases: bug fixing, core implementation, then an optimization phase that “introduces larger inputs that break your phase 2 solution.” Shopify’s interview is “a single open-ended problem that starts simple and evolves.”

Code you didn’t write

Google is reported to be piloting a “code comprehension” round, where candidates analyze an existing codebase with Gemini available, for junior and mid-level roles on select US teams. Stripe’s Bug Bash drops candidates into “an unfamiliar repository with a failing test or open issue.” CoderPad’s survey describes teams “moving away from isolated algorithm puzzles and toward scenarios that mirror day-to-day engineering tasks.”

AI in the open

Some companies stopped policing AI and started grading it. The Register’s headline on Canva: “Thou shalt use AI during interviews.” Meta’s internal announcement, first reported by 404 Media, called the format “more representative of the developer environment that our future employees will work in.” CodeSignal put an assistant inside its IDE and gives reviewers “transcripts of all AI interactions.”

What gets scored is judgment. Canva asks, “Can they identify and fix issues in AI-generated code?” A Meta source quoted by Hello Interview set the bar in one sentence: “Should use AI, but need to show you understand the code.” And this is far from universal. Karat reports that “almost two-thirds of companies still prohibit AI use in interviews.”

A seat in the room

Part of the response is physical. interviewing.io reported in October 2025 that Google’s loop was moving to two virtual interviews followed by three or four in person, and half the FAANG interviewers it surveyed expect their companies to bring in-person rounds back.

Put together, the picture is consistent. Gergely Orosz, who published the survey first in The Pragmatic Engineer, called it “the biggest shakeup of tech interviews in the last 15 years.” The algorithm problem is now one slot in a loop that also asks whether you can hold a codebase in your head, explain a decision, and adapt when the problem changes under you.

Part two

How InterviewClue is built for it

Read those changes as a list of requirements and they describe one specific weakness. The first generation of interview assistants took a screenshot, sent it to a model and printed a reply. One screen in, one answer out, nothing remembered. Every format above targets that loop. The cause of the bug is in a file that isn’t on screen. The follow-up refers to something said ten minutes ago. The requirement you just solved has been replaced.

InterviewClue started from a different premise, the one on our About page: a technical problem rarely fits on one screen. Here is how that design meets each change.

Multi-file projects: context that outlives the screen

InterviewClue builds a session context from the code and requirements you show it. As you move between files it keeps what it has already read, tracks files and their versions, and reconciles overlapping views into one problem state. Regions it couldn’t read with confidence are marked as uncertain rather than filled in. You can inspect the captured context before asking for a solve, and show more of the project when something is missing.

So when the bug is in disputes.py and its cause is a contract in provider.py that you scrolled past five minutes ago, the next answer starts with both.

def recent_status(events):
    newest = events[-1]
    return newest["kind"]

Reasoning, not recall: an answer you can explain

Formats built around “why” punish a block of code with nothing behind it. InterviewClue’s coding workspace puts the reasoning in fixed places: the guarantees the solution has to keep, the approach and why it was chosen, the implementation, time and space complexity with the variables defined, and test cases with the edge cases called out. When the interviewer asks about one part, you go to that part instead of rereading a wall of chat.

Follow-ups work the same way. Ask the question aloud with Discuss and the reply draws on the session context, not a fresh screenshot.

None of this replaces understanding. These interviews are designed to find the candidate who can’t explain the line they just wrote. The workspace is organized so that you can.

Requirements that move: the same shortcut, the current code

In a live session InterviewClue follows the files you view and the edits you make. When the interviewer changes the requirement, you don’t start over and you don’t re-explain. Press Command-Return again, and the solve starts from the code as it now stands and the requirement as it now reads. Sections update in place, so the approach and tests you were discussing don’t scroll away.

Code you didn’t write: the diagnosis first

Debugging rounds reward the explanation more than the rewrite. In Stripe’s Bug Bash, Aced’s guide notes, “a clear, well-articulated diagnosis often carries more than a completed fix.” InterviewClue’s coding responses name the bug by file and line range, explain the cause, and carry the problem’s own constraints, such as code marked do-not-modify, into the guarantees the fix has to keep.

AI-assisted rounds: something to check the AI against

When the interview provides its own assistant, the score comes from verification. Meta lists it among its four criteria, and Canva’s successful candidates “Critically reviewed and improved AI-generated code.” The assistant in the editor writes code. It doesn’t hold the problem. InterviewClue keeps the requirements, the files you’ve read and the edge cases in one place, which gives you something concrete to check generated code against before you accept it.

System design: an architecture you can defend

Design rounds were never solvable by recall, and they carry more weight as typing carries less. InterviewClue’s system-design workspace pairs a diagram with the components, data flow and trade-offs behind it, and lets you trace a single flow through the system when the interviewer asks how a request actually moves.

In-person rounds: preparation, not assistance

No live assistant belongs in a room with a whiteboard, and we won’t pretend otherwise. What carries over is preparation. The Question Bank collects reported interview questions by company, and a snapshot solve turns any of them into a worked answer with the reasoning laid out, so the explanation is yours before you walk in.

What we have and haven’t proven

Everything in this half describes how InterviewClue is designed to work. A feature description is a claim, not a measurement, and that applies to us as much as to anyone else in this market. Our Verification page publishes test reports with their methods, evidence and limits. Today it covers one behavior, global shortcut isolation. Evaluations of multi-file and changing-requirement behavior are not published there yet.