Claude Opus 4.5 squeezed a kernel from 147,734 cycles to 1,487

Anthropic has open-sourced a performance-engineering test task: optimize a kernel running on a simulated VLIW SIMD processor. The baseline takes 147,734 cycles. Claude Opus 4.5 got it down to 1,487, a 99x speedup that beat what most humans managed.

The simulator models a single core that can pack up to 12 scalar ALU operations, 6 vector operations (8 elements each), 2 loads, and 2 stores into every cycle. The baseline kernel issues exactly one operation per cycle, leaving the rest of the machine idle.

The workload is a batched tree traversal with hashing: a tree of height 10 (2047 nodes), batches of 256 elements, 16 rounds, 4096 steps in total. Data gets 1536 words of fast memory, and the 2-loads-per-cycle ceiling binds harder than the compute limit.