Coding Interviews Changed in 2026: The New Formats and How to Pass Them

Something shifted in technical hiring this year that nobody sent a memo about. If you've gone through a software engineering interview recently and found the format unrecognizable — longer rounds, unexpected verbal components, new rules about what tools you can use — that disorientation is real and it has a specific cause.

AI-assisted cheating went from a fringe concern to a documented crisis. A Fabric HQ analysis of nearly 20,000 interviews found that 38.5 percent of candidates showed behavioral signals consistent with AI assistance in early 2026, up from 15 percent just six months earlier. For technical roles specifically, the figure hit 48 percent. Companies that built interview loops assuming occasional cheating are now contending with widespread, systematic cheating — and they rewrote the formats accordingly.

The result isn't a single, coherent response. Different companies took the same problem in opposite directions. Some banned AI outright and hardened their loops against it. Others leaned in and added AI-assisted rounds specifically designed to test judgment rather than recall. Both directions have the same effect on candidates: the preparation that worked before 2026 no longer maps cleanly to what's actually being evaluated.

The Companies That Banned AI — and What Their Interviews Look Like Now

Amazon was among the first major employers to formally ban AI tools during interviews and require candidates to explicitly acknowledge the policy. Goldman Sachs and a handful of other financial services firms locked in similar bans in the first half of 2026. For these interviews, the rules are clear: no AI assistance, and proctoring to detect it.

What "detection" looks like is more sophisticated than most candidates expect. It isn't looking for screen captures or network traffic from an overlay app. Behavioral analysis covers gaze direction (reading off-screen differs from organic memory recall), response latency paired with answer structure (AI-assisted responses arrive after a longer pause but land unusually complete and well-organized), and microphone monitoring for secondary audio input. Candidates who assume that a discreet tool running off-screen is undetectable are operating on an assumption that's behind the current state of detection systems.

For these interviews, the underlying difficulty also went up. The logic is straightforward: if most candidates in the applicant pool have access to AI solutions for standard LeetCode problems, the standard LeetCode problem no longer differentiates. Companies responded by introducing questions with more nuanced edge cases, longer problem statements with deliberate ambiguity, and follow-up modifications that test whether the candidate understands their own solution or memorized a pattern.

The bar also crept upward across levels. Mid-level engineering interviews at several companies now include system design rounds that were previously reserved for senior candidates. Senior engineering loops include staff-level scope questions — architectural decisions involving cost, rollout safety, and partial-failure handling — alongside the algorithmic components. The candidate who prepared for the 2024 version of the loop is showing up to a different examination.

The Companies That Went the Other Direction

Meta began piloting an AI-assisted coding format in late 2025. In these rounds, candidates have access to a choice of AI models — including multiple versions of GPT, Claude, Gemini, and Llama — and are explicitly permitted to use them. What changed is what's being evaluated.

In a traditional coding round, the question is whether you can solve this problem. In an AI-assisted round, the question is whether you can use AI effectively: whether your prompts are precise, whether you catch errors in the model's output, whether you understand what the generated code is actually doing and can explain it, and whether your judgment about when to trust the tool versus override it is sound. A candidate who pastes the problem description into a chat window, accepts the first answer, and can't explain what it does fails. A candidate who uses the tool as a thinking partner, corrects its mistakes, and narrates the decision-making throughout is demonstrating something more valuable than the ability to recall a binary search implementation.

Google introduced a "code comprehension" round that gives candidates access to Gemini as an assistant while asking them to analyze an existing codebase. The round isn't testing whether you can write code under pressure — it's testing whether you can read and reason about code that already exists, a skill that maps closely to the actual work of a senior engineer at a large company. Google interviewers now formally evaluate "AI fluency" as a dimension: prompt engineering, output validation, and the ability to identify where a model went wrong.

These are genuinely new skills. They weren't being tested before 2025 in any systematic way. The candidate who has been grinding LeetCode in isolation, without ever practicing how to use AI as a collaborative tool rather than an answer machine, will find these rounds harder than the ban-and-proctor style.

The Narration Premium

Across both approaches — ban and AI-assisted alike — there's a change in how answers are scored that matters more than any specific format change.

Interviewers at multiple companies have described the same pattern: a candidate who solves a hard problem silently now scores lower than a candidate who solves a medium problem with clear, continuous narration. The value of visible reasoning has gone up relative to the value of correct answers alone, because an AI can produce a correct answer without any reasoning, and interviewers need signals that can't be autocompleted.

This means the internal monologue that most engineers run while solving a problem — the why before the code — needs to come out loud during the interview. Not a running commentary on every keystroke, but explicit narration at decision points: why this data structure instead of that one, what edge cases you're considering before you write the solution, where the trade-off between time and space sits for this specific constraint. The candidate who does this naturally is demonstrating problem ownership. The candidate who solves quietly and narrates only after the fact is demonstrating that they have the answer, which any AI also has.

The practical implication: if you've been grinding problems but not practicing speaking while you solve, you're preparing the wrong skill for the current evaluation.

What Preparation Looks Like Now

The old prep model — solve 150 LeetCode problems, review system design patterns, know Big O — is still necessary. It's no longer sufficient.

Practice narrating your reasoning before your hands move. Before writing a single line of code, explain out loud what the problem is actually asking, what constraints matter, and what approach you're considering and why. Make this a habit during every practice problem. Interviewers notice within the first two minutes whether this is a natural part of how a candidate thinks or a performance triggered by the room.

Learn to use AI as a tool, not an oracle. For companies running AI-assisted rounds, spend time practicing prompt engineering specifically for technical problems. Get used to identifying errors in AI output. Try having an AI model explain its solution back to you and then finding the mistake — this is close to what Meta's format actually tests.

Study edge cases deliberately. The harder questions being introduced in 2026 loops are harder specifically because they contain non-obvious edge cases. Practice identifying edge cases before touching the core logic: null inputs, empty arrays, single-element inputs, overflow, concurrent modification. This is the dimension the new questions are probing.

Know your own solutions. If you practice using AI tools to work through problems, make sure you can explain every line of the output without prompting. The ability to narrate someone else's code is a distinct skill from the ability to write it, and it's being tested directly.

Prepare for the verbal component of system design. The trend toward verbal-format system design rounds — Netflix reportedly runs these without any shared diagramming tool — means you need to be able to articulate architectural trade-offs in plain language under time pressure. Practice explaining design decisions to someone who isn't looking at a whiteboard with you.

If you're preparing for the behavioral component of a technical loop — which has also grown in weight as companies try to offset the AI-assisted noise in technical signals — Meeting Copilot's interview assistant gives you a live briefing during the interview itself. You load your resume and the job description; the assistant surfaces relevant examples and framing when the conversation moves to experience questions, without requiring you to keep those details in working memory while you're also solving a technical problem.

The Loop Is Different, But the Stakes Are the Same

The specific changes — in-person rounds, AI-assisted formats, narration scoring, bar creep — are all responses to the same underlying reality: companies can no longer use the 2022 interview loop as a reliable signal for the 2026 engineering role. They're rebuilding the loop in real time, with different companies landing in different places.

For candidates, that uncertainty is its own kind of challenge. You don't know whether you're walking into a ban-and-proctor room or an AI-permitted room until you're briefed before the interview. Preparing for both — which means building genuine technical fluency and learning to demonstrate it verbally, rather than preparing to recall specific solutions — is the approach that survives either format.

The bar went up because AI made the old bar easy to clear without the underlying capability. Passing the new bar requires showing the thing AI can't show: how you actually think.