ok so the moment i knew my prep was wrong, i was sitting in my apartment with like four hundred leetcode problems done and my recruiter sends me Meta's updated interview page. I read the format section and my stomach dropped. Meta changed their coding onsite in October 2025 and i had been grinding the old way for three weeks straight, sliding window templates, binary tree traversals, timing myself on mediums. All that mattered for the phone screen still but one of the two coding rounds at the onsite is now sixty minutes in CoderPad with a chat panel where you can talk to GPT-5 or Claude Sonnet or Gemini or Meta's own Llama 4 Maverick. They hand you the AI. Three weeks of prep and maybe half of it was relevant to the round that actually decides things, i could have screamed.
Marcus, my buddy who did his Meta loop in January, called me from his car right after his onsite and he was amped. Three panel CoderPad, file explorer on the left, editor in the middle, AI chat on the right. Not a blank function signature, not a single leetcode-style problem. A whole project with classes and data models already written, bug fixes first, then build a new feature, then optimize the whole thing. Three phases within one codebase that keeps growing. Marcus has been through Google, Amazon, two unicorns, and he told me this round felt more like work than any interview he has ever sat through. I believed him because Marcus does not hype things up lol.
The models you can pick from are GPT-4o mini, GPT-5, Claude Sonnet 4 and 4.5, Claude Haiku, Gemini 2.5 Pro, Llama 4 Maverick. Marcus said he swapped between Claude Sonnet for the hard parts and GPT-4o mini for boilerplate and it worked well. The chat can see all files in the project but cannot edit your code, you write or paste everything yourself. Python, Java, C++, C#, Kotlin, TypeScript are the options.
The part that nobody warned me about, Meta changed how they score. Four dimensions now, problem solving, code quality, verification, communication. Verification is new and it is the one people bomb. You have to prove you are reading what the AI spits out before you paste it into your editor. One of Meta's own engineers said "should use AI but need to show you understand the code, explain the output, test before using, do not prompt your way out of it." A guy in my prep group pasted Claude's answer for a graph traversal without checking it, missed a boundary condition in the loop, interviewer flagged it immediately. Done. Did not advance. I watched him do it in a mock the week before too and told him to slow down but he did not listen, classic.
They also want you talking constantly which murdered me at first. My Google loop was code in silence for forty minutes then explain at the end. Meta wants narration from minute one, why you picked this approach, what the AI gave you, whether you agree with it or not. Marcus had to literally snap his fingers through Zoom during our mocks because i kept going quiet for two or three minutes while reading code. Took me four practice sessions to break that habit.
Prep mistake that cost me two full days, i did open-ended distributed system design before learning those are E5 and above. E3 and E4 get API design and client-server stuff. Two days completely wasted because i did not ask my recruiter what level i was interviewing at. Do that before you prep a single hour of design.
Behavioral round is still there, Meta uses CAR not STAR. Context Action Result instead of Situation Task Action Result. I had five stories from Amazon prep in STAR format and converting them took two evenings i did not have. A good CAR result has numbers in it, "cut release cycles from three weeks to four days" works, "improved the process" does not. Have six ready with real metrics before the recruiter call because Meta moves fast. My gap between phone screen and onsite was ten days and some people get less.
i ran InterviewMan during mock Meta sessions, it picked up the questions through audio and read CoderPad in real time, suggestions showed up on screen in about two seconds. During behavioral it pulled a story framework from earlier that matched what my practice partner was asking about which honestly spooked me a little with how well it tracked context. Neither mock partner saw anything on screen share or in the recording after. Works with CoderPad for both the phone screen and AI onsite round. Twelve bucks a month annual. Interview Coder costs two ninety nine monthly for coding only, no behavioral no design. Meta throws all of it at you in one day and paying less than a spotify subscription to cover every round type felt like getting away with something.
The old five hundred leetcode grind still matters for the phone screen. But the onsite is now about reading code you did not write, verifying it under pressure, and talking through your reasoning the whole time. If you are doing a Meta loop in 2026 and you have not drilled that exact combination of skills you are walking in unprepared for the round that matters most.
Ready to Ace Your Next Interview?
Join 57,000+ professionals using InterviewMan to get real-time AI assistance during their interviews.
