Leaderboard

Agentic Rankings

Autonomous tool use and computer operation: OSWorld, Toolathlon, MCP Atlas, τ²-Bench Retail and Apex Agents.

Overall738 Coding437 Math56 Reasoning451 Agentic55 Knowledge450 Multimodal420 Coding Agents55 Scientific Reasoning82 Multilingual12 Writing61 Chat & Assistants61 Research85 Tool Calling32 Long Context82 Speed440 Value439 Open Weights173

55 models

#	Model	Podium Score	OSWorld	Toolathlon	MCP Atlas	τ²-Bench Retail	Apex Agents	Speed
01	Kimi K3 Moonshot AI · Open · reasoning	100.0	—	73.2%	84.2%	—	37.6%	37 t/s
02	Claude Fable 5 Anthropic · reasoning	96.2	—	—	—	98.5%	—	71 t/s
03	Gemini 3.5 Flash Google	75.9	—	—	—	95.3%	—	267 t/s
4	Claude Opus 4.8 Anthropic	74.9	—	59.9%	82.2%	94.4%	—	50 t/s
5	Claude Opus 4.7 Anthropic	74.3	—	—	77.3%	—	—	80 t/s
6	GPT-6 Astra OpenAI · reasoning	72.6	72.6%	—	—	88%	—	82 t/s
7	Qwen3.7 Max Alibaba	71.5	—	—	76.4%	94.7%	—	203 t/s
8	MiMo V2.5 Pro Xiaomi · Open	69.0	—	—	—	94.2%	—	65 t/s
9	GLM-5.2 Z.ai · Open	66.9	—	48.2%	76.8%	99.1%	—	164 t/s
10	Qwen3.8 Flash Next Alibaba · Open	63.5	—	—	—	—	—	168 t/s
11	Gemini 3.1 Pro Google	62.6	—	—	69.2%	95.6%	33.5%	136 t/s
12	GPT-5.5 OpenAI	61.1	—	55.6%	75.3%	93.9%	—	60 t/s
13	GLM-5.3 Flash Z.ai · Open	61.0	—	—	—	—	—	182 t/s
14	Qwen3.7 Plus Alibaba	59.0	—	—	73.2%	—	—	54 t/s
15	Granite 4.2 30B Instruct IBM · Open	58.0	—	—	—	—	—	78 t/s
16	Qwen3.6 Plus Alibaba	52.6	—	39.8%	74.1%	97.7%	—	55 t/s
17	Muse Spark Meta	51.9	—	—	—	91.5%	—	80 t/s
18	DeepSeek V4 Pro DeepSeek · Open	49.7	—	51.8%	73.6%	91.2%	—	66 t/s
19	DeepSeek V4.1 Flash DeepSeek · Open · reasoning	48.0	—	—	—	—	—	214 t/s
20	GPT-5.6 Luna OpenAI	43.3	—	53.4%	—	—	—	175 t/s
21	Claude Opus 4.6 Anthropic	43.1	72.7%	—	62.7%	84.8%	—	50 t/s
22	Kimi K2.7 Code Moonshot AI · Open	43.0	—	—	—	90.1%	—	39 t/s
23	Granite 4.2 8B Instruct IBM · Open	42.0	—	—	—	—	—	115 t/s
24	GPT-5.4 OpenAI	41.6	—	54.6%	67.2%	—	—	60 t/s
25	Agnes 3.0 Flash Agnes · reasoning	40.0	—	—	—	—	—	235 t/s
26	Agnes 2.5 Pro Beta Agnes	37.0	—	—	—	—	—	65 t/s
27	Hunyuan Hy3 Tencent	36.7	—	48.5%	79.1%	—	25.6%	69 t/s
28	DeepSeek V4 Flash Vision DeepSeek	36.0	—	—	—	—	—	120 t/s
29	G9v3-39A5B AI9Stars · Open	36.0	—	—	—	—	—	60 t/s
30	G9v3-3B AI9Stars · Open	36.0	—	—	—	—	—	60 t/s
31	GLM-5.3 Zhipu AI	36.0	—	—	—	—	—	90 t/s
32	Grok 4.6 (low) xAI	36.0	—	—	—	—	—	59 t/s
33	Grok 4.6 (medium) xAI	36.0	—	—	—	—	—	59 t/s
34	Grok 4.6 (xhigh) xAI	36.0	—	—	—	—	—	62 t/s
35	HyperNova 60B Multiverse · Open	36.0	—	—	—	—	—	367 t/s
36	LFM2.5-2.6B Liquid AI · Open	36.0	—	—	—	—	—	60 t/s
37	LFM2.5-8B-A1B Liquid AI · Open	36.0	—	—	—	—	—	338 t/s
38	LongCat 2.0 LongCat · Open	36.0	—	—	—	—	—	42 t/s
39	Magistral Small 1.2 Mistral · Open	36.0	—	—	—	—	—	60 t/s
40	Qwen3.8 27B (xhigh) Alibaba · Open	36.0	—	—	—	—	—	53 t/s
41	Qwen3.8 27B (low) Alibaba · Open	36.0	—	—	—	—	—	64 t/s
42	Qwen3.8 27B (medium) Alibaba · Open	36.0	—	—	—	—	—	62 t/s
43	Qwen3.8 27B (Non-reasoning) Alibaba · Open · reasoning	36.0	—	—	—	—	—	62 t/s
44	MiniMax M3 MiniMax	35.4	—	—	—	88.9%	—	93 t/s
45	DeepSeek V4 Flash DeepSeek · Open	35.3	—	47.8%	69%	—	—	113 t/s
46	GPT-5.6 Sol OpenAI	33.9	—	58%	—	85.1%	—	72 t/s
47	GPT-5.6 Terra OpenAI	30.7	—	53.1%	—	86.3%	—	127 t/s
48	Kimi K2.6 Moonshot AI · Open	26.3	—	50%	—	—	27.9%	40 t/s
49	Gemini 3 Pro Google	24.1	—	—	—	87.1%	—	120 t/s
50	GPT-5.2 OpenAI	17.4	—	46.3%	60.6%	—	—	90 t/s
51	GPT-5.3 Codex OpenAI	17.1	—	—	—	86%	—	118 t/s
52	Gemini 3 Flash Google	15.9	—	49.4%	57.4%	—	—	120 t/s
53	Claude Sonnet 4.6 Anthropic	7.3	72.5%	—	61.3%	—	—	50 t/s
54	Qwen3.5 397B-A17B Alibaba · Open	0.0	—	38.3%	—	—	—	71 t/s
55	Nemotron 3 Ultra 550B NVIDIA · Open	0.0	—	—	—	83.3%	—	142 t/s

Podium Score is a weighted blend of five leaderboards. See Methodology for details.