NanoGPT Speedrun Frontier

NanoGPT Speedrun Frontier NanoGPT 竞速前沿

---

Collapse all models | All models | Best validated result for each model | All trajectories | Log scale | Filter by model 折叠所有模型 | 所有模型 | 每个模型的最佳验证结果 | 所有轨迹 | 对数刻度 | 按模型筛选

---

The following table tracks the performance of various AI models in the “NanoGPT Speedrun” benchmark, detailing their record scores, percentage of gap closed within 24 hours, and operational metrics. 下表追踪了各种 AI 模型在“NanoGPT 竞速”基准测试中的表现，详细列出了它们的记录分数、24 小时内的差距缩小百分比以及运行指标。

Rank	Model	Status	Record	Gap Closed @24H	Agent	Total Tokens	Days
1	Fable 5	note	2,726	81.7%	claude-code · high	800M	8.7
2	Opus 5	serial era	2,920	53.6%	claude-code · max	183M	2.9
3	Kimi K3	serial era	2,930	52.2%	prime-agent · max	112M	3.6
4	Kimi K3	note	2,974	45.8%	kimi-code · max	682M	5.1
5	Opus 4.8	-	3,018	39.4%	claude-code · max	318M	3.0
6	GPT-5.6 Sol	note	3,042	35.9%	codex · xhigh	2.9B	6.1
7	GPT-5.6 Sol Pro	serial era	3,058	33.6%	codex · xhigh	1.2B	3.4
8	Sonnet 5	note	3,105	26.8%	claude-code · max	998M	2.0
9	GPT-5.6 Luna	note	3,110	26.1%	codex · xhigh	894M	1.9
10	Grok 4.5	note	3,120	24.6%	grok-cli · xhigh	46M	2.7
11	Qwen3.8 Max	running	3,120	24.6%	qwen-code · max	216M	1.9
12	GLM 5.2	-	3,150	20.3%	pi · high	57M	1.8
13	DeepSeek V4 Pro	running	3,205	12.3%	claude-code · max	26M	1.1
14	GPT-5.6 Terra	serial era	3,214	11.0%	codex · xhigh	417M	1.1
15	Grok 4.6	running	3,220	10.1%	grok-cli · xhigh	27M	0.6
16	Muse Spark 1.2	running	3,230	8.7%	muse-code · xhigh	41M	0.6
17	Muse Spark 1.1	-	3,232	8.4%	pi · max	122M	3.7
18	GPT-5.5	serial era	3,234	8.1%	codex · xhigh	70M	1.1
19	Kimi K2.7	note	3,240	7.2%	kimi-code · max	160M	1.6
20	GLM 5.3	running	—	no record	claude-code · xhigh	—	—

---

Compare Results: This section provides a comparative view of model records against the Human baseline (2,600) and the target baseline (3,290). 结果对比：本节提供了模型记录与人类基准（2,600）及目标基准（3,290）的对比视图。

Loading available trajectories… Open Traces to explore 41 curated full agent trajectories, including tool calls, subagents, and scratchpads. 正在加载可用轨迹……打开“追踪”（Traces）以探索 41 个精选的完整智能体轨迹，包括工具调用、子智能体和草稿板内容。