Anthropic's 1,579-Cycle Take-Home: The New AI-Resistant Interview

Anthropic's performance-optimization team spent 2024 and 2025 watching its own model beat its own take-home. In January 2026, team lead Tristan Hume published the postmortem. If you are preparing for a technical interview right now, the interesting part is not that Claude cheats. The interesting part is what Anthropic built to replace the questions Claude kept solving, because every serious engineering org is copying it.

What Anthropic actually changed

Anthropic did not ban AI. It redesigned the questions so using Claude stopped helping. AI use is expressly permitted on the take-home; the interview simply no longer rewards it.

Hume's post, "Designing AI-resistant technical evaluations," walks through three iterations of a performance-engineering take-home. The arc:

1. Version 1: A conventional hardware-optimization challenge. Claude Opus 4 outperformed most human applicants.
2. Version 2: Hume rewrote the test to be less about optimizing hardware and more about novel reasoning, sufficient to stump contemporary AI tools, briefly.
3. Version 3: Opus 4.5 matched the best humans on version 2, hitting 1,579 cycles after two focused hours, versus 1,487 cycles after 11.5 hours on the older format. Anthropic then shipped a Zachtronics-style logic puzzle with a tiny constrained instruction set and zero initial debugging tools. Candidates have to build their own tooling before they can even measure their solution.

Anthropic also cut the time limit from four hours to two, and published the original challenge on GitHub inviting anyone to "best Opus 4.5." The point is not secrecy. The point is that when the substrate is unfamiliar enough, the bottleneck becomes the candidate's model of the system, not their retrieval of prior art.

The 48% number reshaping every interview loop

Roughly one in two technical interviews is now flagged for AI-assisted cheating, and the jump happened in a single quarter. That is the pressure making companies rewrite questions.

Fabric, an interview-integrity vendor whose dataset of 19,368 interviews is now cited across the industry, published the following:

Two things to notice. First, 61% of flagged cheaters would have passed anyway. The interview was already easy enough that they did not need help; they took it because a $20 to $50 monthly Cluely subscription against a $150,000 salary is a risk-reward ratio that heavily favors cheating. Second, the tech-vs-sales gap is not about ethics. It is about what the questions measure. Sales interviews test judgment in a live conversation, which is genuinely hard to feed through an overlay. LeetCode-style screens are trivial to feed.

The new question shape, and how to prepare

The 2026 technical interview rewards taste, learning speed on an unfamiliar system, and the ability to interrogate model output rather than accept it. That is the shape you should prepare for.

Anthropic's Zachtronics-style format is the clearest example, but the pattern is showing up everywhere. Here is what to actually practice.

1. Build-your-own-tools problems

Expect a constrained instruction set, no standard library, and no debugger. Your first hour is spent writing the print statements and profilers you will need to solve the actual problem. Claude can generate a print statement, but it cannot decide which invariant is worth instrumenting on your specific mental model of the system.

2. "Prompt an AI, then defend its output"

Several companies now let you use Claude or GPT during the interview, then grade you on how you challenge its answer. A junior engineer accepts the model's output as truth. A senior engineer interrogates it, challenges its assumptions, and forces it to refine against constraints the model did not know existed.

To rehearse this, take a real system you own, ask Claude to optimize a hot path, and write down every load-bearing assumption in its response before you run it. If you cannot list five, you are not ready to be interviewed on this format.

3. Live system-model interviews

Instead of a whiteboard problem, you get a real codebase you have never seen and a specific question about it. The interviewer watches how you navigate: what you grep for, which files you open first, whether you run the tests. AI helps here, but only if you already know what to ask it.

4. In-person and attested formats

Google's Sundar Pichai addressed AI cheating at an internal town hall, and the company is considering reintroducing mandatory in-person interviews for certain roles. Amazon now requires candidates to sign an attestation acknowledging unauthorized-tool policies before assessments. If a company flies you out, assume the entire loop is now the interview, including lunch.

Getting to that in-person round starts with a resume the recruiter actually reads. That is the work Refolk takes off you: paste the posting, get your resume back rewritten for it, with a cover letter drafted and a fit score you can trust before you spend a Saturday on their take-home.

Why juniors are getting squeezed hardest

The redesign taxes junior engineers twice: they cheat at roughly double the rate of seniors, and the new questions specifically measure the taste juniors have not yet accumulated. That is not fair, but it is the market.

Fabric's data shows 0-to-5-year candidates cheating at nearly 2x the rate of senior candidates. The economic pressure is obvious: junior loops are longer, rejection rates are higher, and the ROI on an invisible overlay looks great until you get flagged. But even the ones who do not cheat lose ground when the interview format shifts from "implement this algorithm" (which juniors can grind for) to "here is Claude's answer, tell me what it got wrong" (which requires having been burned by a subtle bug in production at least once).

The counter is boring and it works:

• Ship something real that has users, and read every bug report yourself.
• When you use Claude, keep a log of the times it was confidently wrong. Bring three examples to every interview.
• Contribute to an open-source project with a maintainer who will review your PRs harshly. Public code review is the fastest taste accelerant available.

The AI-fluency credential almost nobody has yet

Publicly credible multi-tool AI fluency is still vanishingly rare, which means it still functions as a differentiator on a resume in early 2026.

In Refolk's index of professional profiles, roughly 518 US software engineers publicly list both Claude and GitHub Copilot as skills. Against a pool of 212,681 senior and staff SWEs in the same index, that is about 0.24%. The concentration is extreme: Microsoft alone accounts for 10 of the top-25 sample, with Meta and Intuit at 2 each.

Two implications for your resume:

1. Listing "AI tools" is not a credential. Everyone claims it. Listing specific tools, with specific outcomes ("cut a Rust build's incremental compile from 47s to 9s using Claude-assisted profiling"), is still rare enough to move a recruiter.
2. The signal degrades fast. Use it now, before the line item is table stakes.

If you want the tailored version of that pitch on every application without rewriting your resume 40 times, that is what Refolk does: it reads the posting, pulls the parts of your history that map to it, and writes the resume and cover letter to match. The version you send to Anthropic emphasizes the profiling story; the version you send to a fintech emphasizes the ledger-consistency story, without you doing that work twice.

The tools companies are actively detecting

If you are considering using a live-assist tool, know exactly what interview platforms are looking for. This is the current detection meta.

The named tools Fabric and comparable platforms explicitly detect:

• Cluely, rebranded from "cheat on everything" to a general meeting assistant in November 2025 without changing its underlying invisible overlay
• Interview Coder
• Final Round AI
• Leetcode Wizard
• Parakeet AI

The base rate is what it is: 48% of technical interviews flagged, 61% of those flagged would have passed anyway. Do the math on your own risk tolerance, and remember that 83% of candidates say they would use live AI if they thought they would not be caught. Detection vendors are pricing against that number, not against the honest minority.

What to actually do this week

Rebuild your prep around three specific artifacts, not around LeetCode reps.

1. A "Claude was wrong" doc. Three to five specific bugs, with the prompt, the output, and the fix. This is your answer to the "defend the AI's output" question, and it lands in interviews.
2. One build-your-own-tools project. Pick a constrained instruction set and write the interpreter, then the debugger. You are training for the exact shape of Anthropic's new take-home.
3. A resume that names specific tools and specific outcomes. Not "AI-native." Not "prompt engineering." The line item that separates you from 99.76% of the senior SWE pool is "Claude + Copilot, with numbers." Refolk will draft that version from your actual history and re-tailor it for every posting you send.

The recruiter reading your application is now handling more than twice as many candidates as in 2021 and conducting about 40% more interviews per hire, according to Gem's Recruiting Benchmarks. They are exhausted, and they are looking for reasons to say no. Give them a reason to say yes that Claude did not write for them.

FAQ

Is it actually against the rules to use AI on a take-home in 2026?

It depends on the company, and increasingly the answer is no. Anthropic expressly permits Claude on its performance-optimization take-home; the questions are just designed so that using it does not help. Amazon requires an attestation before the assessment that specifies which tools are and are not allowed. Read the instructions on the assessment page word by word, and if the policy is silent, ask the recruiter in writing. A permitted use documented in email is worth ten times a permitted use you assumed.

How do I practice for an AI-resistant interview when I have never seen one?

Work in constrained instruction sets where Claude has no docs to lean on. Rebuild an interpreter for a small fictional assembly language, or take an unfamiliar codebase and give yourself two hours to answer a specific question about it without opening the model. The reflex you are training is "build the tool you need to see the problem," which is exactly what Anthropic's new format measures. Anthropic also published its original challenge on GitHub, so you can attempt it directly.

Will listing Claude and Copilot on my resume actually help, or is it noise?

It helps right now because it is still rare. In Refolk's index, only about 518 US software engineers, roughly 0.24% of the senior pool, publicly list both. Make it specific. "Claude + Copilot" alone is noise. "Reduced p99 latency on our billing service from 340ms to 90ms using Claude-assisted flame-graph analysis" is a credential.

If 48% of interviews are flagged, are recruiters just assuming everyone cheats?

Some are, which is why the honest-candidate false-positive problem is real. The defense is to make your process legible. Narrate your thinking on live calls, keep a visible scratchpad, share your screen proactively when solving, and reference specific past bugs by name. Interviewers flag opacity, not competence. The candidates who get through the 48% flag rate are the ones whose work is easy to watch.