Anthropic gave its own AI the take-home it gives engineering candidates. Claude did better than most of the humans.

Then Anthropic made the test harder. A newer Claude matched the best human score within two hours.

So here is the question every hiring team should ask this week. If the company that builds the model can't keep a coding test ahead of it, how long will yours last?

The technical interview for AI engineers changed more in 2026 than in the ten years before it. The labs say openly that their AI solves the tests they used to give. Hiring teams across Europe are inventing their own formats. These include logic tests, challenges deployed to the cloud, design rounds where the interviewer interrupts you and regulation questions no US loop would ask.

This post is long on purpose. It's a reference you can come back to, whether you're a candidate preparing for a loop or, like most of our readers, a hiring team deciding what your own interview should look like.

We start with how Anthropic and OpenAI interview engineers as of November 2026. That includes the real tasks candidates have reported and one take-home Anthropic published itself. Then we zoom out to Europe, drawing on an excellent 20-market study by AI Engineering Insider. We close with what I think it all means for anyone designing a technical interview today.

A note on evidence. Neither Anthropic nor OpenAI publishes its full question bank. Where a question below comes from candidate reports rather than an official source, I say so.

How Anthropic Interviews Engineers

Anthropic's software engineering loop is practical, fast and unusually heavy on values. The company doesn't publish one fixed sequence of rounds, so the outline below combines Anthropic's own careers guidance with consistent candidate reports from 2026.

The Typical Loop

Stage What happens Typical length
Recruiter call Background, motivation, "why Anthropic" and interest in AI safety 15–30 min
Online assessment A timed CodeSignal project that grows in levels About 90 min
Technical screen Live, practical coding with an engineer 50–60 min
Hiring manager Experience, judgment and past technical decisions 45–60 min
Virtual onsite Coding, system design and a culture round, often split over two days 4–6 sessions
References and team match References, then matching to a team Can take weeks

Interviews run remotely over Google Meet. CodeSignal and Google Colab are used for the technical work. Candidates report timelines from about four weeks to three months, mostly depending on team matching and reference checks.

What Makes It Different

The problems are practical. Candidates describe coding rounds as close to real work. You build a small working system, then extend it as the requirements change. Interviewers ask many follow-up questions to see whether your design survives the new requirements. Speed matters too. Several candidates say the hardest part is getting real tests passing on a multi-part problem within the hour.

Concurrency shows up everywhere. Candidates for infrastructure teams in particular report crawlers, async code, thread-safe caches and timeouts again and again.

System design is shaped by Anthropic's products. Infrastructure teams get inference and model-distribution problems. Product teams get product-flavoured designs, such as a prompt playground.

The culture round is the real gate. This is a conversation about AI safety, ethics, judgment, mistakes and disagreement. interviewing.io, drawing on conversations with Anthropic engineers, reports that this values round is where most candidates fail. Recruiters send culture and values documents beforehand. Candidates are expected to have formed their own view of them, including where they disagree.

The AI Rule

Anthropic publishes official guidance on candidates' AI use. The short version:

  • Application. Write your own first draft, then refine it with Claude.
  • Preparation. Use Claude freely to research the company and practise answers.
  • Take-homes and live interviews. No AI assistance unless Anthropic explicitly says otherwise.

The guidance also lists what Anthropic itself uses Claude for in hiring. That includes writing job descriptions, developing interview questions and transcribing interviews.

The "unless we say otherwise" part is growing. Candidates describe a 55-minute AI-assisted round in some loops since mid-2026. It comes with a browser editor, a repository with four pull requests and Claude Code.

One Manager Dropped the Coding Interview

On 27 October 2026, Felix Rieseberg, who leads engineering for Claude.ai, Claude Code Desktop and Cowork, wrote that the coding interview is dead for his team. He had removed it from his loops that month. Claude is now better than candidates at small coding challenges, so a quick React exercise tells him about as much as an assembly quiz would.

He now looks at CVs and portfolios for small prototypes that show "taste, empathy, and principled thinking". He also invited other managers to share better methods, because nobody has fully solved what replaces the coding round.

Two caveats matter here. This is a decision for one team, and Anthropic hasn't announced any change for the whole company. Its careers page still lists live coding tools for technical roles. And Anthropic's official candidate guidance hasn't changed. There is still no AI in live interviews unless you're told otherwise.

That tension is the story. The company with arguably the strongest coding model bans AI in most interviews, allows it in some and has dropped coding rounds altogether on at least one team.

How OpenAI Interviews Engineers

OpenAI is one of the few AI labs that writes down its own process in a public interview guide. The guide covers every role, so it describes the shape of the loop and leaves the details to each team.

The Typical Loop

Stage What happens Typical length
Application review The recruiting team reviews your CV About a week
Introductory calls A recruiter and often the hiring manager ask about background, motivation and team fit About 30 min each
Skills-based assessment Pair coding, a take-home project or a technical test, and some roles have more than one Varies by team
Final interviews 4–6 hours with 4–6 people over one or two days, virtual by default with an onsite option in San Francisco 1–2 days
Decision Usually about a week after the finals, and references may be requested About a week

For software engineers, candidate reports fill in the middle. The screen is often a one-hour coding round plus a one-hour system design round, sometimes back to back. The final day usually adds more coding, more system design, a project presentation or deep dive and behavioural interviews on collaboration and mission. The whole process usually takes between two and eight weeks.

What OpenAI Says It Grades

OpenAI's stated engineering bar is explicit. It looks for well-designed solutions, high-quality code, optimal performance and good test coverage, plus communication and collaboration (DataInterview's summary of the guide).

Test coverage on that list is worth noticing. Many candidates prepare for algorithms and forget that tests are part of the grade.

Candidates describe the coding rounds as heavy on implementation. One widely shared comment says they were typing for the whole hour and passed the last test case just in time. You choose your language, and the questions are matched to it.

The AI Rule Depends on the Round

OpenAI's guide says that some interview formats intentionally allow AI and other tools. Others are designed to test your independent problem-solving without them. The prep materials for each interview say which applies, and candidates are told to ask their recruiter if they're unsure.

In practice, candidate guides describe most coding rounds as AI-free, with one notable exception. OpenAI is piloting an agentic coding round in the onsite. You get a prompt, an existing codebase and a problem that is deliberately too large to solve by writing code from scratch. The expectation is that you work through it with an AI coding agent. The round is still in beta, so not every candidate gets it (interviewing.io).

Anthropic and OpenAI Side by Side

Anthropic OpenAI
Published process High-level careers guidance plus a detailed AI-use policy A public interview guide that covers every role
Default AI rule No AI in take-homes or live interviews unless stated Varies by format, and the prep materials say which
Rounds that allow AI Some AI-assisted rounds and the performance take-home An agentic coding round in beta
Coding style Practical builds in several levels, heavy on concurrency Practical builds with a lot of implementation, and tests are graded
Distinctive round The culture and AI safety interview, often decisive A project presentation or deep dive
Recent shift One manager dropped coding interviews for his team (October 2026) An agentic coding round in pilot

The common direction is clear. Both labs have moved from LeetCode-style puzzles to realistic engineering tasks. Both are visibly experimenting with where AI belongs in the interview.

Real Tasks From Anthropic and OpenAI Interviews

One Anthropic take-home is fully public, published by Anthropic itself. Everything else in this section comes from candidate reports collected by prep sites and forums. Neither company has confirmed these questions, and the exact wording varies. But the same problems are reported again and again.

Anthropic's Performance Engineering Take-Home

In January 2026, Anthropic open-sourced its original performance take-home on GitHub. You get a Python simulation of a custom accelerator with manually managed scratch memory, and a kernel that works but is slow. Your job is to rewrite it so the same computation finishes in as few clock cycles as possible.

The story behind it is the best case study I know on what AI does to technical assessments. In Designing AI-resistant technical evaluations, Tristan Hume, a lead on Anthropic's performance optimisation team, describes three versions of the test.

  1. Version 1 was a four-hour take-home from early 2024. Claude Opus 4 eventually did better than most human applicants given the same time limit.
  2. Version 2 was a harder two-hour version that started from faster code. Claude Opus 4.5 then matched the best human performance within two hours.
  3. Version 3 is a new take-home designed to resist AI. Anthropic says it is curious how long it will last.

Two details are worth noting for anyone who designs assessments. This take-home explicitly allows AI use, unlike Anthropic's default policy. And the public repository still works as a recruiting channel. Anthropic invites anyone who beats its published model benchmark to email their code to the performance recruiting team.

Reported Anthropic Coding Tasks

  • The in-memory database in levels. This is the most common online assessment. Level 1 is basic get, set and delete. Later levels add search and filtering, then expiry based on a time to live, then queries over historical values. Some versions end with file compression and surviving a restart. Candidates repeat one lesson. Write modular code from level 1, or you'll be refactoring under time pressure at level 3.
  • The web crawler. Build a crawler that stays within one domain and removes duplicate URLs, then make it multithreaded. Interviewers push on concurrency, timeouts, redirects and relative versus absolute URLs more than on pure optimality.
  • From profiler samples to a trace. Convert snapshots from a sampling profiler into an ordered timeline of start and end events, handling recursion and repeated stacks. A variant asks you to find the slowest function.
  • An image transformation pipeline. Apply a series of image operations described in JSON, sometimes with a library you haven't used before.
  • Finding duplicate files. Find the duplicates in a very large directory tree without loading whole files into memory.
  • An in-memory banking service. Accounts and transfers with timestamped operations and careful handling of edge cases.
  • A tokenizer. Implement longest-match tokenization over a vocabulary, with follow-ups on counting tokens and pricing input and output tokens.
  • System design prompts. Design Claude's chat service, a prompt playground, batched inference or a system that distributes models.

Reported OpenAI Coding Tasks

One poster on Blind summarised OpenAI's recent question bank as eight recurring problems. They are a cd command, a spreadsheet, an in-memory database, a key-value store, a resumable iterator, node counting, GPU credits and a dependency version check. Other prep sites report the same family.

  • Serialising a key-value store. Serialise and deserialise a store where keys and values can contain any character, including your delimiter. Most strong answers land on length-prefix encoding, the pattern the Redis protocol uses.
  • A time-based key-value store. Every write carries a timestamp, and reads return the value active at a given time. Follow-ups cover locking and persistence.
  • The cd command. Resolve paths with ., .., absolute paths and symbolic links, including cycle detection.
  • A spreadsheet engine. getCell and setCell with formulas, a dependency graph and detection of circular references.
  • Resumable iterators. Iterators over nested structures that can pause, resume, skip and reset.
  • GPU credit management. Credits that expire over time on a first-in, first-out basis, with add, expire and consume operations.
  • An in-memory SQL-like database. Create tables, then insert, delete and select with conditions.
  • A multithreaded web crawler. Yes, both labs ask it.

What the Two Lists Have in Common

There are no trick puzzles and no obscure graph algorithms. Almost every task is a small, realistic system with state, edge cases and a follow-up that changes the requirements.

That is much closer to everyday engineering. It is also much harder to fake with memorised LeetCode answers.

And, as the take-home story shows, it is exactly the kind of task current AI models handle well. That is why both labs are now experimenting with where AI belongs in the loop.

Europe Has Twenty Different AI Interviews

If you hire AI engineers in Europe, the US labs are only part of the picture. On 3 November 2026, the newsletter AI Engineering Insider published AI Engineering Interviews in Europe: 20 Markets, 150 Questions. It's the most thorough public map of European AI interviews I have seen. The team collected candidate reports, job postings and labour-market data from the UK to Estonia, and kept only questions sourced from 2024 to 2026. Their headline finding is simple. Europe does not have one AI interview.

What follows is my summary and commentary. For the full list of 150 questions, the detail for each market and their paid guide with model answers, read the original.

What Every Market Shares

Across all twenty markets, the technical core is the same. Candidates are expected to understand attention, fine-tuning techniques such as LoRA and preference tuning, how to evaluate a RAG system, agents that use tools and the cost of inference. The study also found that most European AI jobs are applied LLM work rather than model pre-training. Postings at companies such as Zalando, Booking.com, UiPath and Feedzai ask for agents, MCP, RAG and evaluation.

Two trends in the study stood out to me as an interviewer.

Evaluation is now a job in itself. Booking.com builds evaluation as an internal service using LLM judges. Corti sets clinical quality gates before a model can ship. DeepL runs multilingual human evaluation. If your interview never asks candidates how they would know a model is good enough, you're missing a core skill.

Regulation is part of system design. This is the layer no US loop has. The EU AI Act's Article 50 transparency rules apply from August 2026, and the Annex III high-risk obligations have moved to December 2027. GDPR shapes every data decision. Interviewers increasingly ask about HR, credit and health use cases directly. They expect regulation to appear in the design answer from the start.

Where European Loops Differ

The study's most useful insight for hiring teams is that loops differ more in format than in topic. Here are some examples, summarised from the article.

Company City What makes the format distinctive
Google DeepMind London Two coding rounds, an ML fundamentals round, ML design, then a lead researcher interview
Bending Spoons Milan An online logic and probability test before any human conversation, then homework and a proctored exam
ML6 Ghent A challenge you deploy to a public cloud project before the first call, then a paper presentation
Adyen Amsterdam A 30-minute ML design round where the interviewer deliberately interrupts you
Synthesia London A GPU take-home on a decoder language model
Booking.com Amsterdam Two business cases, one on the spot and one prepared, such as payment fraud
Stripe Dublin An ML "bug squash" round where the training script runs but the model is worse than it should be
Mistral AI Paris A technical HR call, deep questions on transformer and inference internals, and questions on how you work with coding agents

Several of these formats are clever answers to the problem the US labs face. A challenge deployed to the cloud, a bug-squash round and a design session with deliberate interruptions all test things a candidate can't easily hand to an AI model during the interview. They test judgment, debugging when the problem is unclear and composure.

Demand and Pay Are Uneven

The study draws on Indeed's AI Tracker for demand. In July 2026, the share of job postings mentioning AI was highest in Ireland at 15%, followed by Spain at 13% and the UK at 9%. Germany had the most AI-labelled job titles, and most of them were outside tech occupations.

Pay varies even more. Zurich pays the highest base salaries. Polish postings often publish their salary ranges. Packages at the frontier labs in London sit far above typical market medians.

For a hiring team this matters. A loop that is long, unusual or exam-like costs you more candidates in a market where they have many options.

Fifteen European Interview Questions

I picked fifteen of the study's 150 questions that show the range. They're paraphrased in my own words and grouped by what they test. The company, role and year come from the original, which also marks each question as reported by candidates, attributed to a vendor or derived from job postings. Treat the topics as the signal and the wording as approximate.

ML fundamentals, in depth

  1. Explain the KV cache, estimate its memory cost per token and say how grouped-query and sliding-window attention change it. (Mistral AI, 2025)
  2. Compare PPO and GRPO for training a language model to reason. (Google DeepMind, London, 2026)
  3. Derive the gradient of cross-entropy through a softmax using the chain rule. (DeepL, 2026)
  4. A full fine-tune of a 7B model runs out of memory on eight 80 GB GPUs. Walk through the memory budget and fix it. (Mistral AI, 2024)

Debugging and judgment

  1. Your training loss suddenly spikes. Separately, inference is too slow for the product. What do you do in each case? (Google DeepMind, London, 2024)
  2. A training script runs without errors, but the model is worse than expected. Find the bugs. (Stripe, 2025)
  3. There's no time to run an experiment before a decision. What do you do? (Wolt, 2025)

Evaluation

  1. How do you evaluate a RAG system? (Mistral AI, 2026)
  2. A support agent claims to resolve half of all conversations. How would you measure the resolution rate honestly? (Intercom, 2026)
  3. Design the evaluation for an LLM that writes clinical notes, and decide when a new model may ship. (Corti, 2026)

System design for real products

  1. Design a system that generates short recaps of newly published podcast episodes. (Spotify, 2026)
  2. Design an internal MCP server that gives agents secure access to many data stores. (Riskified, Lisbon, 2026)

Regulation and language in Europe

  1. A user invokes the GDPR right to erasure. How do you remove their data from a RAG index, logs, analytics and a fine-tuned model? (Core question, 2026)
  2. A public-sector customer requires that no data leaves the EU and that the model can run on their premises. Design the deployment. (Core question, 2026)
  3. Your assistant must work in all 24 official EU languages. How do you evaluate it without testing only in English? (Core question, 2026)

One more question deserves a mention of its own. Mistral AI reportedly asked a software engineer in 2026 when they last wrote code without help from an AI agent. It's a short question with a lot underneath it, and I expect to hear versions of it in many more loops.

For the other 135 questions, including an overview of each of the twenty countries, go to the original article.

What Hiring Teams Should Take From This

The best-funded AI companies in the world are still figuring out how to interview engineers now that AI writes code. That's reassuring for everyone else. There is no settled answer you're failing to copy.

To be fair to the coding interview, it isn't dead everywhere. One manager dropped it for one team, and both labs still run live coding in most loops. What has changed is that a coding round on its own no longer tells you enough.

There are clear lessons, and they line up with how we design structured interviews at AssessDev.

1. Decide What Each Round Measures, Then Decide the AI Rule

Anthropic bans AI by default. OpenAI decides round by round. One Anthropic manager has dropped the coding round entirely. These look contradictory, but they share one principle. The AI rule follows from what the round is meant to measure.

If a round tests raw problem-solving, AI hides the signal. If it tests how someone directs and checks an AI agent in a large codebase, AI is the point. Write down the signal for every round first. The AI rule then becomes obvious, and you can tell candidates clearly in advance.

2. Replace Puzzles With Small Systems That Change

The in-memory database in levels, the crawler that becomes concurrent and the key-value store that gains timestamps all work for the same reason. Each follow-up changes the requirements. You see how a candidate structures code, handles edge cases and adapts. That predicts far more than whether they have memorised a graph algorithm. It also gives every interviewer a consistent rubric, level by level.

3. Test Judgment, Debugging and Evaluation as Well as Building

The strongest European formats are a step ahead here. Stripe's bug-squash round, Corti's question on whether to ship and Wolt's question about having no time for an experiment all test judgment under uncertainty. These skills matter more when AI writes much of the code. Someone still has to notice that the numbers are wrong.

4. Treat the Values Round as a Real Assessment

At Anthropic, the culture round reportedly decides more outcomes than any coding round. Whatever your company's values are, a values or collaboration interview needs the same rigour as a technical one. That means defined signals, consistent questions and trained interviewers. A friendly chat at the end of the day is not an assessment.

5. In Europe, Put Regulation and Language Into System Design

If you hire AI engineers for European products, your system design prompts should include the AI Act, GDPR, EU data residency and multilingual evaluation. A candidate who designs a brilliant RAG system with no answer for the right to erasure is not ready for a European product team.

6. Keep Your Loop Proportionate to Your Market

A logic test before any human contact, or a cloud challenge before the first call, can work for a company that candidates already want to join. Most teams compete for engineers in markets like Ireland, Spain or Zurich. For them, every extra hour of unpaid assessment costs candidates. Measure your drop-off at each stage, and cut what doesn't add signal.

7. Train the Interviewers

Every format above depends on the person running it. An interrupted design round only works if the interviewer interrupts consistently. A coding task in levels only works if every interviewer scores level 3 the same way. Most broken loops we see have good questions. Untrained interviewers use them inconsistently.

Three Things to Do This Week

Write one line per round. Next to each round in your loop, write the one skill it measures and whether AI is allowed. If you can't fill in a line, that round needs work.

Add one follow-up to your coding task. Pick the step where the requirements change, and write down what a weak, solid and strong answer to it looks like.

Ask one judgment question. Add a debugging or "would you ship this?" question to your next interview, and agree on the scoring before the candidate joins.

So, honestly: if Claude sat your coding round tomorrow, would it pass?

Rethinking Your Technical Interview for 2027?

This is the work we do at AssessDev. We train interviewers, run structured technical interviews for you and embed a CTO-level interviewer in your team, with a method built for how engineering works now.

Is Interview as a Service Right for You?

Not sure where to start? Our free Buyer's Guide to Interview as a Service explains when handing your technical interviews to a partner makes sense, and when it doesn't.

Sources

The European section of this post is based on AI Engineering Interviews in Europe: 20 Markets, 150 Questions by AI Engineering Insider, published 3 November 2026. All European findings, companies and example questions come from their research. The summaries and commentary are mine. Their full 171-page guide includes model answers for all 150 questions, tables of each company's loop, an overview of tech stacks, an eight-week plan and pay data for all twenty markets. If you're preparing for or running AI interviews in Europe, subscribe to their newsletter.

Official sources

Reporting and guides based on candidate reports

Candidate-reported questions are not confirmed by Anthropic or OpenAI. Interview processes change often, so check each company's official guidance before you interview.

Photo by Pranjall Kumar on Unsplash.