NanoGPT Speedrun Frontier: leaderboard
NanoGPT Speedrun Frontier evaluates autonomous coding agents on the nanoGPT optimizer speedrun task. The model must improve the training recipe on its own under different agent harnesses. The pinned metric is human record gap closed, measuring the share of the distance between the baseline and human record that the run closes. Source: BenchmarkList’s attributed result table.
Metric: Human record gap closed (%). Source: benchmarklist.com. Saturation forecast: Around February 2027. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5 (High) | 81.7 | #20 (Claude Fable 5) |
| 2 | Claude Opus 5 (Max) | 53.6 | #16 (Claude Opus 5) |
| 3 | Kimi K3 | 52.2 | #35 |
| 4 | Kimi K3 (Max) | 45.8 | #35 (Kimi K3) |
| 5 | Claude Opus 4.8 (Max) | 39.4 | #39 (Claude Opus 4.8) |
| 6 | GPT-5.6 Sol (xHigh) | 35.9 | #22 (GPT-5.6 Sol) |
| 7 | Claude Sonnet 5 (Max) | 26.8 | #65 (Claude Sonnet 5) |
| 8 | GPT-5.6 Luna (xHigh) | 26.1 | #79 (GPT-5.6 Luna) |
| 9 | Qwen 3.8 Max (Max) | 24.6 | #34 (Qwen 3.8 Max) |
| 10 | GLM-5.2 (High) | 20.3 | #80 (GLM-5.2) |
| 11 | DeepSeek V4 Pro (Max) | 12.3 | #112 (DeepSeek V4 Pro) |
| 12 | GPT-5.6 Terra (xHigh) | 11 | #40 (GPT-5.6 Terra) |
| 13 | Muse Spark 1.2 (xHigh) | 8.7 | #69 (Muse Spark 1.2) |
| 14 | GPT-5.5 (xHigh) | 8.1 | #33 (GPT-5.5) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=nanogpt-speedrun-frontier · How It Works · Data refreshed daily, snapshot 2026-11-05.