Since October 2025, one of the two coding rounds at a Meta onsite hands you something that would have gotten you walked out of the building a year earlier: a chat panel with GPT, Claude, Gemini, and Llama sitting in it, and explicit permission to use any of them. Google started piloting its own version in May 2026, Gemini-only, for junior and mid-level engineers on Google Cloud and the Platforms and Devices team. Neither company made the interview easier. They made it harder in a way most candidates are preparing for backwards.
The reflex is to assume the AI does the work now, so the bar dropped. It didn’t. Once the assistant can write a correct sliding-window solution in four seconds, writing one yourself stops being the thing worth measuring. What they measure instead is whether you can tell a good answer from a plausible-looking wrong one, and whether you can steer the model to the right place faster than the candidate in the next slot.
What the screen actually looks like
Meta’s round runs in CoderPad with three panels: a file explorer on the left, an editor in the middle, and an AI chat on the right. You get a multi-file codebase, not a blank function stub. The assistant replies in the chat but cannot touch your files, so every suggestion it makes is something you have to read, judge, and paste in yourself. You can switch models mid-session, so a common move is to ask Claude Sonnet for a design and cross-check a tricky edge case against Gemini. You have sixty minutes. The question arrives in stages that tend to run bug fix, then implement a feature, then optimize, all inside the same codebase and theme. A representative prompt: here’s a small service with a red test suite, get it green, then add cursor-based pagination to the results endpoint.
Google’s pilot, which they call a code comprehension round, drops you into an unfamiliar codebase of roughly 200 to 500 lines. You have an hour to read it, find the bugs, explain design choices someone else made, and propose improvements, with Gemini available the entire time. The backdrop matters: Sundar Pichai said in April 2026 that 75% of new code at Google is now AI-generated and approved by an engineer, up from about half the previous fall. The interview is starting to resemble the job.
| Dimension | Meta AI-enabled coding round | Google code comprehension round |
|---|---|---|
| Rollout | Live since October 2025, replaces one of two onsite coding rounds | Pilot from May 2026, junior/mid SWE, US teams (Cloud, Platforms & Devices) |
| AI models offered | GPT, Claude Sonnet, Claude Haiku, Gemini, Llama; switchable mid-round | Gemini only |
| Starting material | Multi-file codebase, one theme across stages | Unfamiliar codebase, ~200-500 lines |
| Duration | 60 minutes | 60 minutes |
| Task arc | Fix bugs, implement a feature, optimize | Read, locate bugs, justify design, suggest improvements |
| Can the AI edit your files? | No, chat suggestions only; you paste | Chat assistant alongside the editor |
| Scored on | Problem solving, code quality, verification, communication | Prompt engineering, output validation, debugging |
| Classic no-AI round still present? | Yes, one traditional algorithm round remains | Standard DSA rounds unchanged so far |
The four things Meta is actually scoring
Meta grades this round on problem solving, code quality, verification, and communication. Google names its three as prompt engineering, output validation, and debugging. Different vocabulary, same center of gravity: can you use the tool without being used by it. The two that decide most outcomes are verification and communication, and both are the ones candidates neglect because they feel like soft skills next to writing the code.
Verification is the whole game. When you paste in an AI suggestion, the interviewer is watching what you do next. Do you run it? Do you add a test for the case the model conveniently ignored? Do you read the diff, or skim it because it looks right? The senior habit of distrusting output you didn’t check is exactly the signal they built this round to surface.
The candidates who fail trust the model
The most common way to lose is to accept the assistant’s first answer because it compiles and reads with confidence. Say a test is failing, you ask the model to fix it, and it hands back code that passes by special-casing that one input while quietly breaking the general path. If you paste it and move on, you just demonstrated the anti-pattern the whole format exists to catch. The interviewer saw it. That single move sinks the verification score no matter how tidy the code looks.
The failure mode has a tell. Watch for the moment the model produces something longer or cleverer than you expected, and you feel relief instead of suspicion. That relief is where people stop reading. Strong candidates get more skeptical exactly when the answer looks good, because a confident wrong answer costs more than an obvious one. Gemini and GPT will both state incorrect complexity claims in a calm, authoritative voice, and part of the round is noticing when the confident tone and the actual code disagree.
What a good prompt looks like
A vague prompt wastes your most limited resource, which is minutes. Compare these two.
Weak: fix this bug
Strong: merge_ranges() returns overlapping intervals when the
input isn't pre-sorted. It's called from schedule() at
line 88, which assumes the output is disjoint. Give me a
fix that sorts defensively inside merge_ranges, keep it
O(n log n), and flag anything in schedule() that relied
on the old behavior.The strong prompt names the function, the symptom, the caller, the constraint, and asks the model to surface downstream effects. It reads like a well-written ticket. That is not an accident of style; the interviewer is reading your prompts as a work sample. Precise prompts that isolate a problem area, hand over context, and ask for focused help are what score high on Google’s prompt-engineering axis, and they happen to be the prompts that actually get you a usable answer.
The other habit worth building is asking the model to explain before it writes. A prompt like walk me through what this function assumes about its inputs before you change anything forces the assistant to expose its reasoning, which gives you something to verify against instead of a black-box patch. It also keeps you talking, which feeds the communication score.
How to practice without gaming it
You cannot cram this with a problem set the way you can grind array questions. The skill is reading code you didn’t write and staying skeptical of a confident assistant, and both take reps. Pull a mid-size open-source repo in a language you know, break something on purpose, and fix it with an AI assistant open while you narrate every decision out loud. Time-box it to an hour so you feel the same pressure. The goal is to get fast at the read-judge-paste-verify loop, not at typing.
Practice catching the model being wrong. Ask it something you already know cold, then watch how it fails and how quickly you notice. Ask it for the time complexity of code you wrote and check whether it’s right. The muscle you’re building is the reflex to treat every output as a claim that needs evidence, not an answer. Do this enough and the suspicion becomes automatic, which is the point.
Keep narrating even while you read the chat panel. Long silent stretches while you scan the model’s reply read as passivity to an interviewer who can’t see inside your head. Say what you’re checking for, say when the answer looks wrong and why, say when you’re switching models and what you expect the second one to catch. The people who clear this round sound like an engineer reviewing a junior’s pull request, not someone hoping the machine got it right.
None of this replaces fundamentals. Meta still runs one classic round with no AI at all, and you still need to recognize when a problem wants a heap or a topological sort, because you can’t verify an answer you don’t understand. The shift is that knowing the algorithm is now the floor, not the ceiling. The engineers who do well in 2026 are the ones who were already reading AI output with a raised eyebrow at their day job, and it shows within the first ten minutes.
Practice the behavioral round:
