Benchmark profile
KernelBench Hard H100 (KernelBench)
An agentic GPU-kernel benchmark that measures how much of the hardware roofline a model's correct, audit-clean kernels reach on six demanding CUDA and Triton problems.
How we show KernelBench
We mirror the KernelBench Hard H100 PCIe board from kernelbench.com. The visible score is mean peak fraction of roofline (valid cells) across 6 GPU-kernel problems; failed and audit-flagged cells do not enter that mean.
This is an independent agent benchmark, not the Stanford KernelBench project leaderboard. Rows combine a model, an agent harness, long-running GPU sessions, and human audit decisions, so the table stays display only.
Mean peak fraction of roofline on KernelBench — July 17, 2026 snapshot
BenchLM mirrors the published mean peak fraction of roofline view for KernelBench. Claude Fable 5 leads the public snapshot at 23.9% , followed by Grok 4.5 (22.9%) and Kimi K3 (20.9%). BenchLM does not use these results to rank models overall.
Claude Fable 5
Anthropic
claude-fable-5
Grok 4.5
xAI
grok-4.5
Kimi K3
Moonshot AI
kinetic-0715
Mean peak fraction of roofline table (14 models)
ScoreThe published KernelBench snapshot is tightly clustered at the top: Claude Fable 5 sits at 23.9%, while the third row is only 2.9 points behind. The broader top-10 spread is 13.5 points, so the benchmark still separates strong models even when the leaders cluster.
14 models have been evaluated on KernelBench. The benchmark falls in the Coding category. This category carries a 20% weight in BenchLM.ai's overall scoring system. KernelBench is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About KernelBench
Year
2026
Tasks
6 GPU-kernel optimization problems
Format
Mean peak fraction of hardware roofline over valid cells
Difficulty
Agentic GPU systems engineering
The mirrored board comes from the independent kernelbench.com Hard suite, not the Stanford KernelBench project. One long-running agent session tackles each problem. The visible score averages peak fraction of roofline over valid cells; failed and audit-flagged cells are excluded from that mean. We keep the benchmark display only because harness, hardware, and audit state are inseparable from the score.
BenchLM freshness & provenance
Version
KernelBench 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
FAQ
What does KernelBench measure?
An agentic GPU-kernel benchmark that measures how much of the hardware roofline a model's correct, audit-clean kernels reach on six demanding CUDA and Triton problems.
Which model leads the published KernelBench snapshot?
Claude Fable 5 currently leads the published KernelBench snapshot with 23.9% mean peak fraction of roofline. BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on KernelBench?
14 AI models are included in BenchLM's mirrored KernelBench snapshot, based on the public leaderboard captured on July 17, 2026 snapshot.
The AI models change fast. We track them for you.
A weekly brief for engineers and researchers covering new models, ranking shifts, and pricing changes.
Free. No spam. Unsubscribe anytime.