순위표

도구 호출 랭킹

에이전트 워크플로의 함수 호출 및 API 오케스트레이션: Toolathlon, MCP Atlas, τ²-Bench.

종합700 코딩399 수학53 추론413 에이전트55 지식412 멀티모달382 코딩 에이전트55 과학 추론82 다국어12 글쓰기61 채팅 어시스턴트61 리서치85 도구 호출32 긴 컨텍스트82 속도402 가성비401 오픈 웨이트170

32개 모델

#	모델명	Podium Score	Toolathlon	MCP Atlas	τ²-Bench Retail	속도
01	Claude Fable 5 Anthropic · reasoning	98.5	—	—	98.5%	71 t/s
02	Gemini 3.5 Flash Google	95.3	—	—	95.3%	267 t/s
03	MiMo V2.5 Pro Xiaomi · 오픈소스	94.2	—	—	94.2%	65 t/s
4	Muse Spark Meta	91.5	—	—	91.5%	80 t/s
5	Kimi K2.7 Code Moonshot AI · 오픈소스	90.1	—	—	90.1%	39 t/s
6	MiniMax M3 MiniMax	88.9	—	—	88.9%	93 t/s
7	GPT-6 Astra OpenAI · reasoning	88.0	—	—	88%	82 t/s
8	Gemini 3 Pro Google	87.1	—	—	87.1%	120 t/s
9	GPT-5.3 Codex OpenAI	86.0	—	—	86%	118 t/s
10	Qwen3.7 Max Alibaba	85.6	—	76.4%	94.7%	203 t/s
11	Nemotron 3 Ultra 550B NVIDIA · 오픈소스	83.3	—	—	83.3%	142 t/s
12	Gemini 3.1 Pro Google	82.4	—	69.2%	95.6%	136 t/s
13	Claude Opus 4.8 Anthropic	78.8	59.9%	82.2%	94.4%	50 t/s
14	Kimi K3 Moonshot AI · 오픈소스 · reasoning	78.7	73.2%	84.2%	—	37 t/s
15	Claude Opus 4.7 Anthropic	77.3	—	77.3%	—	80 t/s
16	GPT-5.5 OpenAI	74.9	55.6%	75.3%	93.9%	60 t/s
17	GLM-5.2 Z.ai · 오픈소스	74.7	48.2%	76.8%	99.1%	164 t/s
18	Claude Opus 4.6 Anthropic	73.8	—	62.7%	84.8%	50 t/s
19	Qwen3.7 Plus Alibaba	73.2	—	73.2%	—	54 t/s
20	DeepSeek V4 Pro DeepSeek · 오픈소스	72.2	51.8%	73.6%	91.2%	66 t/s
21	GPT-5.6 Sol OpenAI	71.6	58%	—	85.1%	72 t/s
22	Qwen3.6 Plus Alibaba	70.5	39.8%	74.1%	97.7%	55 t/s
23	GPT-5.6 Terra OpenAI	69.7	53.1%	—	86.3%	127 t/s
24	Hunyuan Hy3 Tencent	63.8	48.5%	79.1%	—	69 t/s
25	Claude Sonnet 4.6 Anthropic	61.3	—	61.3%	—	50 t/s
26	GPT-5.4 OpenAI	60.9	54.6%	67.2%	—	60 t/s
27	DeepSeek V4 Flash DeepSeek · 오픈소스	58.4	47.8%	69%	—	113 t/s
28	GPT-5.2 OpenAI	53.5	46.3%	60.6%	—	90 t/s
29	GPT-5.6 Luna OpenAI	53.4	53.4%	—	—	175 t/s
30	Gemini 3 Flash Google	53.4	49.4%	57.4%	—	120 t/s
31	Kimi K2.6 Moonshot AI · 오픈소스	50.0	50%	—	—	40 t/s
32	Qwen3.5 397B-A17B Alibaba · 오픈소스	38.3	38.3%	—	—	71 t/s

Podium Score는 5개 리더보드의 가중 평균입니다. 자세한 내용은 평가 방법을 참고하세요.