Anthropic Rewrote Its Take-Home Test Three Times: Claude Opus 4.5 Caught Up With the Best Engineers

Anthropic has used its performance engineering take-home test since early 2024. More than 1000 candidates have taken it, and dozens of them now work at the company, including people who brought up the Trainium cluster and shipped every model since Claude 3 Opus. Now every new version of Claude breaks the test. At the same time limit, Claude Opus 4 outperformed most humans. Claude Opus 4.5 matched the best candidates, and at the two-hour mark their results could no longer be told apart from the model's. The test's author, Tristan Hume, a lead on the performance optimization team, described this on January 21, 2026.

How the test was built

In November 2023, Anthropic was preparing to launch Claude Opus 3. TPU and GPU clusters were already in place, a large Trainium cluster was expected, and the company was short on performance engineers. Hume posted on Twitter, and more candidates wrote in than the standard pipeline could handle. He spent two weeks designing the test.

He built a Python simulator of a fictional accelerator modeled on the TPU. Candidates optimize code on it while watching a hot-reloading Perfetto trace that shows every instruction. The machine has manually managed scratchpad memory, VLIW (multiple execution units per cycle), SIMD and multicore. The task is deliberately not about deep learning: it is a parallel tree traversal inspired by branch-free SIMD inference of decision trees. Candidates start from a fully sequential implementation, then do a multicore warm-up and choose between SIMD and VLIW packing. The first version still contained a bug that candidates had to debug, which meant building a tool for the job.

The very first result confirmed the idea. A candidate from the Twitter wave scored noticeably higher than the rest, joined the team, and almost immediately found a workaround for a compiler bug that was blocking a release: tensor index math was overflowing 32 bits.

How Claude broke the test

By May 2025, more than half of candidates would have done better simply by handing the task to Claude 3.7 Sonnet through Claude Code. The pre-release Opus 4 produced a solution better than almost all humans within the 4-hour limit. Hume used the model to find the point where it starts to stall and made that the starting point of the second version. He wrote cleaner starter code, added new machine features, removed multicore, which Claude already handled, and cut the limit from 4 hours to 2, because scheduling slots was stretching hiring out by weeks.

That did not work against Opus 4.5. The model closed the main bottlenecks, implemented the standard micro-optimizations, and cleared the passing threshold in under an hour. Then it stopped, concluding that it had hit the memory bandwidth limit. Most humans reach the same conclusion. But there is a trick that exploits the structure of the problem. When Hume told Claude the achievable cycle count, it thought about it and found the trick. By the 2-hour mark its result equaled the best human result, and that human was himself a heavy user of Claude 4. In an internal harness with test-time compute, the model kept improving its result as time increased.

Two attempts to save the format

Hume rejected banning AI: it cannot be enforced, and it moves the test away from real work. The option of "beating Claude Code on your own" is also questionable. Claude works fast, while humans spend half their time reading the problem statement, so the winning strategy could become watching from the sidelines.

The first new task grew out of Hume's own optimization work: transposing data in TPU 2D registers without memory bank conflicts. Opus 4.5 came up with something the author had not anticipated: transposing not the data but the whole computation. Hume closed that loophole. The model fell short of the optimum, but with ultrathink mode and an increased thinking budget it still solved the task. The reason is that a lot has been written about bank conflicts, and Claude has a huge store of experience to draw on.

The second attempt went in a weirder direction. Hume needed a task outside the training data distribution, so he turned to Zachtronics games, which use unusual, tightly constrained instruction sets. This conflicts with the goal of making the test look like real work, and the author acknowledges the trade-off.

Key facts

• The test has been in use since early 2024, and more than 1000 candidates have taken it.
• The time limit was cut from 4 to 2 hours.
• Claude 3.7 Sonnet outperformed more than 50% of candidates.
• Opus 4 beat almost all humans in 4 hours.
• Opus 4.5 cleared the threshold in under an hour and matched the best human in 2 hours.
• With unlimited time, the best humans still outperform Claude.

Practical takeaways for hiring teams

The model solves tasks built on well-known techniques, such as transposition or bank conflicts, by drawing on accumulated experience. It is more reliable to test reasoning on unfamiliar material than to test knowledge. Banning AI achieves little: it is easier to design the test on the assumption that the candidate works with a model. The job itself has also shifted toward complex debugging, systems design, performance analysis and correctness verification, and these are hard to assess in a couple of hours. Anthropic has published the original assignment as an open challenge: anyone who beats the Opus 4.5 result can write to the company.