Claude solves problems better than candidates, and an Anthropic manager gave up coding interviews
Plenty of developers spend weeks grinding LeetCode problems just to get through a single round of technical interviews. One Anthropic manager has decided that round no longer makes sense – Claude beats most candidates at those puzzles anyway.
That manager is Felix Rieseberg, an engineering manager at Anthropic who has led development of Claude.ai and the Claude Code desktop app. Earlier this week, Rieseberg posted on X that the live coding interview had stopped telling him anything useful about candidates. He pulled it from his team's hiring process in October.
He explained the decision in more detail on his blog. He used to watch a candidate define a problem and arrive at a neat solution – a weak but real predictor of how they'd perform on the team.
Once Claude got good at those exercises, a small React challenge started revealing about as much about a person as a quiz on assembly.
Anthropic's performance team ran into the same problem. Team lead Tristan Hume wrote in January that Claude Opus 4 outscored most applicants on the take-home test within the time limit. Hume made the test harder, but Opus 4.5 arrived later and matched even the strongest candidates.
The third version of the test now relies on unusual, puzzle-game-style problems. Opus 4.5 couldn't crack those, so the format remains a workable way for the team to separate human skill from model capability.
Rieseberg hasn't found a full replacement for the interview yet. For now, he's going through resumes and portfolios looking for small prototypes that show taste, empathy for users, and principled thinking. He also asked other managers in the industry to share better evaluation methods.
Google principal engineer Jaana Dogan was among the first to reply. In her response on X, she wrote that she'd rather watch candidates work with coding agents, since that process reveals more about their real skills.
Most of Rieseberg's essay is aimed at junior engineers, and its title, "Hunger beats tenure," sums up the main idea. He argues that newer engineers can now outpace veterans because building working products has become so cheap.
As an example, he points to two fairly new Anthropic engineers who built a working graphical interface for Claude Code over a single weekend. The first public version shipped about eight days later.
The broader mood around AI and new grads is a lot gloomier, though hiring data is still mixed. ZipRecruiter labor economist Nicole Bachaud told The Register that tech employers are still the most likely to test candidates' skills directly.
ZipRecruiter's October survey covered 1,500 employers. Among those in the information sector, 62.5 percent require formal skills tests as part of the initial application process. Around 70 percent have moved those assessments to proctored or in-person sessions to keep AI out.
Anthropic's own candidate guidance still asks applicants to keep Claude out of live interviews and take-home tests unless told otherwise. Rieseberg dropped the coding round only for his own team, so anyone applying elsewhere at Anthropic will likely still face one.
- Vibe coder created a drone-shooting shooter in 8 hours for $550 with one short prompt in Fable 5.2 – waiting for the first AI hit worth millions
- AI enthusiast spent $2,618 on a single prompt and got Rome of 320 AD overnight
- In 44 hours, Claude Fable 5.2 built a playable Gears of War clone from a three-line prompt, but the compute was valued at nearly $15,000