/Puzzles
About
Team
Investments
Research
Research Index
Build
IncubationsOpen Source
Writing
Paradigm Puzzles
Hackathon

Terms, Disclosures, Privacy

LinkedIn, Twitter, Contact

← Back to Leaderboard

About the Anthropic Take-Home Challenge

The Challenge

Anthropic open-sourced its original performance take-home — the optimization test it used before Claude started beating humans at it. You are given a Python-simulated VLIW machine and a working but slow kernel in perf_takehome.py. Your job: rewrite build_kernel so the same computation finishes in as few clock cycles as possible.

This leaderboard began as a community project (kerneloptimization.fun) and now lives in the arena, with all standings preserved.

Community project — not affiliated with or endorsed by Anthropic.

The Machine

The simulator is a wide-issue VLIW core: each cycle it dispatches one instruction bundle across parallel engines — alu, valu, load, store, and flow. Your Python code runs once to build the instruction stream; the machine then executes it. The score is how many cycles that execution takes. Keeping every engine busy every cycle is the whole game: scheduling, vectorization, software pipelining, and algebraic shortcuts all pay off.

Scoring

  • Submit your complete perf_takehome.py. The judge runs the upstream correctness shape: forest_height=10, rounds=16, batch_size=256, output values required, output indices optional.
  • To keep seed-fishing out, that same shape is repeated across nine fresh random seeds. Your official score is the worst (highest) cycle count observed. Lower is better.
  • The leaderboard shows each account's best score; equal scores rank by whoever achieved the score first.

Benchmarks

A handful of model reference runs appear as non-competing rows in the rankings, so you can see where your kernel lands:

  • 999 cycles — Claude Code Fable 5 (max effort, 4-hour one-shot)
  • 1,363 cycles — Claude Opus 4.5 (Anthropic reference)
  • 1,378 cycles — Claude Code Opus 4.8 (max effort, 3-hour one-shot)
  • 1,404 cycles — Opencode GLM 5.2 (iterative, 48 hours)
  • 1,990 cycles — Opencode GLM 5.2 (3-hour one-shot)

These are orientative only — single data points under specific harness conditions, not guaranteed reproducible. Replication attempts may yield different results.

Submission Rules

  • Paste or upload the complete perf_takehome.py (at most 1,000 KB).
  • Each X account can submit 10 times per day; only your best score counts.
  • Execution is sandboxed in two phases: your code runs only to generate instructions, which are serialized, re-sanitized (engine whitelist), and simulated by trusted code. Submitted Python never runs during scoring.
  • Solutions that exploit the harness instead of optimizing the kernel are rejected.

Community

Special thanks to @dougallj, whose vliw-challenge.fly.dev leaderboard kept the challenge alive.

Resources

The take-home itself — simulator, starter kernel, and correctness tests — lives in anthropics/original_performance_takehome. Clone it and run python tests/submission_tests.py to check which thresholds you pass locally before submitting.

LeaderboardSubmitGitHubAbout
Dan RobinsonParadigm
LeaderboardSubmitGitHubAbout
Dan Robinson