Google, Stripe, Meta, and AWS run four different loops in 2026, and the gaps between them have widened in the last twelve months. Round counts diverge. AI-tool policies diverge. What each panel weights most heavily diverges. An engineer optimizing for one loop is, in practice, mis-optimizing for the others. This post lays the four loops next to each other on the same axes — count, hours, AI policy, top-weighted dimension, retake window — so the trade-offs are visible at a glance.
Why this matters
Engineers preparing for one of these loops often prepare for all four, in case offers stack. The plans they build are usually one-size-fits-all: drill algorithms, build a system design portfolio, rehearse behavioural. That plan is fine for 2022. It is not fine for 2026. The four loops now ask measurably different questions, weight different dimensions, and permit different tools in the room. A candidate who lands a Stripe offer using AI in the coding round may stall at a Google onsite where the same tool is disabled. A candidate who built a defensibility habit on a take-home will do better at Meta and AWS, where the panels probe it directly.
What does each company actually weight most heavily?
Weighting is the part recruiters won't tell you and that engineers reverse-engineer from rejection letters. We pulled the signal from 1,440 tracked rounds across these four employers and the 14 others in our interview-fit study. The picture is not subtle.
Google still weights coding correctness highest. Their panel rubric has not changed materially since the 2024 revision — correctness, problem-solving approach, code quality, and communication, in roughly that order. What did change is the threshold. Google now expects the correctness bar to be met without AI assistance, on the assumption that any senior engineer can hit it with AI; the loop's job is to confirm the floor.
Stripe weights defensibility — the candidate's ability to explain why a decision was right — above raw correctness. Stripe's coding round is the only one of the four that gives the candidate a constrained AI tool by default. The trade is explicit: you can use the tool, but the panel will ask you to defend every line that came from it. Candidates who paste AI output without reading it lose the round on defense, not on code.
Meta weights system design more heavily than any of the other three for L5+. The coding round still happens, but the offer or no-offer call is made in the system design and behavioural rounds. Meta panels have started asking for a "decision replay" — a five-minute walkthrough of a recent architectural decision the candidate made with AI in the loop.
AWS weights operational thinking. Their bar raiser explicitly probes failure modes: what happens at 10x scale, what the page-out looks like at 3am, what gets rolled back first. The coding round at AWS is still hard, but it is the supporting evidence, not the verdict.
| Company | Rounds | Total hours | AI policy | Top-weighted | Retake window |
|---|---|---|---|---|---|
| 5 | 5.5 | Disabled | Coding correctness | 12 months | |
| Stripe | 4 | 4.5 | Constrained, scored | Defensibility | 6 months |
| Meta | 5 | 5.0 | Disabled (L5+ replay) | System design | 6 months |
| AWS | 4 | 4.0 | Disabled | Operational thinking | Varies by org |
Which loops permit AI assistance — and where does that leave you?
Stripe is the outlier. Their coding round in 2026 ships with a constrained assistant: a chat panel scoped to the problem, no internet, no clipboard out. The candidate can ask the assistant anything about the problem and the assistant will answer. The panel sees the full transcript afterwards. This is not a generous accommodation. It is a test. The panel scores not whether the candidate used the assistant but how they used it — whether they accepted answers without reading, whether they spotted errors in the suggestion, whether they pushed back when the suggestion was wrong.
Google, Meta, and AWS still disable AI tooling in the room. The reason given by all three is the same: they want a baseline measurement of the candidate without the tool, on the assumption that any modern engineer will have the tool at work. The disabled-room reading is a floor check.
Meta added a wrinkle in Q4 2025. For L5 and above, the system design round now opens with a "decision replay" — the candidate walks the panel through a real architectural decision they made in the last 12 months in which AI played a role. The panel probes which suggestions were accepted, which were rejected, and why. It is not a coding round. It is a defensibility round.
The honest implication: prep that focuses only on raw coding speed without AI will leave you under-prepared for Stripe and partially under-prepared for Meta. Prep that focuses only on AI fluency without baseline coding will leave you under-prepared for Google and AWS. The four loops together require both halves of the skill.
How long do they take, end to end?
In-room hours are only part of the answer. End-to-end calendar time matters more for candidates juggling multiple loops. Google's loop is the longest in-room — five rounds totalling around 5.5 hours — but those rounds get compressed into a single virtual onsite day. AWS's loop, at four rounds and four total hours, is paradoxically the longest calendar-wise, often spread across two weeks because the bar raiser is scheduled separately. Stripe's loop runs about 10 days from screen to offer call. Meta's is closer to 14 days.
The single-day Google format has its own implications. Candidates report that the fifth round at Google is materially weaker than the first — fatigue is real, and the panel knows it. The first two rounds are where most strong-hire signals are generated. The last two are where weak-hire signals are confirmed. The middle round is the swing. A candidate who knows the loop's shape budgets their energy accordingly; a candidate who treats all five rounds as equal usually leaves the fifth one on the table.
The funnel below is a representative loop — it does not perfectly describe any single employer, but it captures the shape candidates can plan against. The compression from initial screen to in-room rounds matters because each compression stage has its own filter. Note how steep the drop is from technical phone screen to in-room — most of the loop's selectivity happens at one stage, not spread evenly across all five.
How do retake and feedback policies differ?
Retake windows used to be uniform — twelve months across the industry. The 2026 picture is messier. Google holds at twelve months for a No Hire outcome at a senior level. Stripe and Meta both shortened to six months for engineers who advanced to onsite but did not receive an offer. AWS varies by org: some teams will re-interview in three months, some hold to twelve, and the candidate has no way to know in advance which org they fell into.
Feedback is the bigger divergence. Stripe gives the most detailed feedback of the four — a written summary covering the panel's rubric scores by dimension, with at least one specific area to work on. Meta gives high-level feedback, usually a single sentence about the dimension that fell short. Google and AWS, by policy, give almost no feedback beyond the verdict. AWS's recruiter will sometimes share a one-line summary off the record; Google's almost never will. The asymmetry matters: the loops that lock you out longest are also the ones that tell you least about why.
The practical implication: if you have a choice in sequencing, run Stripe early. The feedback is the most useful input for the next loop. Run Google last, because the twelve-month lockout is the most expensive miss. Meta and AWS sit in the middle of the order, with Meta slightly ahead because its six-month window matches Stripe's and the system design feedback transfers cleanly to AWS's operational round. Candidates who pick the order by retake window and feedback quality, rather than by which company they want most, finish more loops with offers in hand.
The validation study behind why take-home performance predicts loop outcomes.
From Codritium Research
Take-Home vs. Live Coding, 2026
Does take-home performance predict live-coding outcomes? A longitudinal validation across 412 candidates and 1,860 interview rounds at 23 employers — with a decision tree, calibration plot, and the residuals that broke our first model.
Common questions
Which loop is hardest?
Google has the highest in-room hour count and the strictest coding bar; Meta has the most weight on system design at L5+. There is no single answer — the hardest loop is the one whose top-weighted dimension is your weakest. Use the comparison table to pick.
Do any of these companies use Codritium?
Some engineering teams at the four employers use the platform's challenges and replay format as part of their internal upskilling. We do not power any of their hiring pipelines, and the platform score is not visible to any of their recruiters.
Can I use AI during the interview itself?
Only at Stripe, and only inside a constrained assistant the panel monitors. Google, Meta, and AWS disable AI tooling in the room. Meta's L5+ system design round asks you to discuss a prior decision you made with AI, but the room itself is AI-free.
How long should I prep for each?
Engineers in our tracked cohort who cleared at least one of the four loops averaged 7.4 weeks of focused prep. The 8-week plan in our companion post covers a representative practice arc across all four loops.
Do offers differ by loop tier?
Yes. Each company runs an internal calibration where the panel verdict translates to a level — L4 vs L5 at Google and Meta, SDE II vs SDE III at AWS, IC3 vs IC4 at Stripe. Strong Hire verdicts can shift the level up; Hire verdicts usually map to the level posted on the role.