xAI Grok 4.6 Model Card - KernelBenchInternal v1.1: leaderboard
xAI's internal KernelBench-inspired suite: agents replace PyTorch operators with custom CUDA kernels that must compile and match the reference, scored on correctness and the measured speedup over a TF32 PyTorch baseline. xAI Grok 4.6 model card (revised 2026-08-17), section 5.3 chart; accuracy in percent, higher is better.
Metric: Accuracy (%). Source: media.x.ai. Saturation forecast: Around September 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 (Max) | 48.1 | #16 (Claude Opus 5) |
| 2 | Kimi K3 (Max) | 41 | #35 (Kimi K3) |
| 3 | Claude Opus 4.8 (Max) | 37.5 | #39 (Claude Opus 4.8) |
| 4 | Grok 4.6 (High) | 37.2 | #29 (Grok 4.6) |
| 5 | GPT-5.6 Sol (Max) | 34.4 | #22 (GPT-5.6 Sol) |
| 6 | GPT-5.5 (xHigh) | 34 | #33 (GPT-5.5) |
| 7 | Grok 4.5 (High) | 29.7 | #54 (Grok 4.5) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=xai-grok-4-6-model-card-kernelbenchinternal-v1-1 · How It Works · Data refreshed daily, snapshot 2026-11-05.