Anthropic's CodeSignal Assessment in 2026
The Anthropic CodeSignal screen evolved. Here’s what changed, and how to prep for six.
Anthropic CodeSignal round now has 6 parts, not 4
You sit down for the Anthropic screen expecting four levels. You’ve practiced four levels. Ninety minutes in, you clear what you think is the final stage - and a fifth set of requirements unlocks. Then a sixth.
That happened to one of my clients two weeks ago. She is sharp. She had drilled the in-memory database problem until she could build it in her sleep. She still walked away rattled, because the test she prepared for was not the test she sat.
Some of my clients have now described the same thing in the last month. Anthropic’s CodeSignal screen appears to have grown from four evolving stages to six. Two more waves of requirements you cannot see when you start.
Anthropic does not announce these things, and right now the public interview-prep sites still describe four levels. But when three independent candidates report the identical change in the same window, I pay attention. And if you have a screen coming up, so should you.
Define your data model carefully at the start
Here is why six hurts more than four, and it is not the obvious reason.
The whole point of this assessment was never speed. It is whether your early design can absorb requirements you have not seen yet. With four levels, a shaky Level 1 could limp through to the end. With six, every shortcut you take in the first fifteen minutes compounds twice more. The people who struggle are not the slow coders. They are the ones who built Level 1 to pass Level 1.
My updated advice given the changes to the interview structure is to spend longer up front on the data model, then move fast on features. Designing for change is now the entire game.
Use of AI
The stage count is not the only thing that moved this year. A few changes are well documented, and they matter just as much.
The no-AI rule is real and it is enforced. This is the part people underestimate. Anthropic is an AI company, and they design these screens to resist AI on purpose. They published a piece in January 2026 called “Designing AI-resistant technical evaluations” - the whole philosophy is to test judgment a model cannot shortcut for you. Reaching for an assistant during the screen is not just against the rules. It is detectable, and it can override an otherwise strong score.
The problem bank rotated. The old favourites - the file deduplicator, the in-memory key-value store - still appear, but reviewers are now watching for memorised solutions. If your code reads like you have seen the exact problem before, that is a flag, not a flex.
The grader is a black box by design. The spec is deliberately ambiguous. The strongest candidates do not read the prompt and plan it perfectly in their heads. They write code, run it against the hidden tests, and reverse-engineer what the system actually wants. Treat the failing tests as the real spec.
And recalibrate the scoring in your head. With six stages the points scale with them, and you earn partial credit stage by stage as your code clears the hidden tests, so there is no single magic number to chase. Two things hold regardless of the exact total: a clean design that survives all six stages beats squeezing partial credit out of a brittle one, and the integrity layer can quietly disqualify even a top score. Build for change, not for points.
How to prepare for the revised CodeSignal assessment?
So if you have an Anthropic screen on the calendar, here is what I would actually do this week.
Assume six stages, not four. Budget your ninety minutes as if two more waves are coming, because they probably are.
Spend the first twenty minutes on the data model, not the first feature. Ask yourself one question before you write real logic: if a stranger handed me a requirement I cannot see right now, would my structure bend or break?
Practice by building small systems end to end, not by memorising problems. A key-value store, a file system, a rate limiter. Build each one cold, then bolt on a feature you did not plan for. That is the exact muscle the test measures.
And practice with no AI help at all. If you lean on a model while you prepare, you are training a muscle you will not have in the room.
The candidates who clear this screen are not the ones who memorised the most. They are the ones who design like more is coming. Because now, it is.
If you are prepping for Anthropic - or OpenAI or DeepMind - and you want the full map of how the research-track pipeline fits together, from this first screen through the safety round, I have put all of it into my Anthropic research careers guide. It is the same map I walk my clients through, and it is here: https://www.sundeepteki.org/company-guides#anthropic
Cheers,
Sundeep
Sundeep Teki, PhD
Website | Coaching | LinkedIn | X
P.S. If you have a screen booked and want a second set of eyes on your prep before you sit it, that is exactly what a first coaching call is for.
You can grab a discovery slot here: https://cal.com/sundeep-teki/15min
