AI Agent Tops Kernel Benchmark With Single-Kernel Design, Outpacing Rivals by Wide Margin

An AI agent named Fable has produced the fastest single-kernel submission ever recorded on KernelBench-Mega, a benchmark for evaluating AI-designed GPU kernels. The system wrote CUDA code for an RTX PRO 6000 Blackwell that ran 18.71 times faster than an optimized PyTorch baseline.

That performance blew past rival models. Claude Opus 4.8 managed a 14.4x speedup using Triton code, while GLM-5.2 hit 11.14x and GPT 5.5 reached 4.34x. Fable's approach was also structurally distinct: its solution used exactly one cooperative kernel launch per decoded token, while other top entries split the work across 4 to 14 separate launches.

The result matters because kernel design is a foundational bottleneck in AI development. When AI systems get good at writing optimized kernels, they accelerate the hardware-software stack that trains and runs them. That creates feedback loops: better kernels mean faster training, which can yield better AI systems capable of writing even better kernels. Benchmarks like KernelBench-Mega offer a concrete way to measure how close we are to that kind of recursive self-improvement.