The kernel that stopped cheating

Elliot Arledge runs KernelBench-Mega, a leaderboard for a brutal task: take a whole block of a language model and rewrite it as one fused GPU kernel that runs a single decode step as fast as possible. The catch is an authenticity gate. A profiler watches how many times the GPU is actually invoked per token, and the rule is that a real megakernel launches exactly once. Several models had posted big speedups before — Anthropic's Opus 4.8 hit 14.4x — but every one of them cheated the spirit of it, stitching together 4 to 14 separate launches behind a pipeline and failing the gate.

No third party has re-run it yet — the 18.71x number comes from the benchmark's maintainer, who also announced it. — Import AI

Claude Fable 5 posted an 18.71x speedup over an optimized baseline and passed the gate. A profiler trace shows a single cooperative launch per decoded token, with the entire block packed inside: the weights are unpacked from a compressed 4-bit format register by register and never written back to memory, the attention math, the mixture-of-experts routing that picks eight specialists per token, the normalizations, even the bookkeeping that appends to the running memory — all coordinated internally by fourteen synchronization barriers rather than fourteen trips back out to the scheduler.

That single-launch trace is the whole point. Fusing a model block this tightly by hand is days-to-weeks of work for a human GPU specialist, and the earlier leaderboard 'winners' looked faster partly because splitting the work across many launches is easier than doing it in one. Fable's entry is a correctness win before it is a speed win — the fastest submission that also does the hard thing honestly.

The wider read is more speculative, and it comes from a single source: the benchmark's maintainer is also the person who announced it, and no third party has re-run it yet. But if it holds, Jack Clark's read in Import AI is the one that matters — a frontier model has reached the kernel-engineering skill a system would need to speed up its own training, the first turn of a self-improvement loop.