Agentic · tools

τ²-bench results leaderboard

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap

As of 2026-10-07, the highest measured score on τ²-bench results is 99.1% by GLM-5.2. 55 more models have estimated scores, calibrated from the benchmarks they were measured on.

Measured scores: benchlm.ai.

#	Model	Score	Source
1	BTL-3	100.0%	estimated ± 14.1 pp, low confidence
2	Claude Opus 5.5	99.9%	estimated ± 4.8 pp, low confidence
3	Muse Spark 1.3	99.9%	estimated ± 4.8 pp, low confidence
4	Grok 4.7	99.9%	estimated ± 4.8 pp, low confidence
5	GLM-5.2	99.1%	measured
6	GPT-5.4	98.9%	measured
7	Claude Fable 5	98.5%	measured
8	GLM-5-Turbo	98.5%	measured
9	GLM-5V-Turbo	98.5%	measured
10	Step 3.7 Flash	98.5%	measured
11	GLM-5	98.2%	measured
12	GPT-5.5	98.0%	measured
13	GLM-5.1	97.7%	measured
14	Grok 4.3	97.7%	measured
15	Qwen3.6 Plus	97.7%	measured
16	DeepSeek V4 Pro 0813	96.2%	measured
17	Grok 4.6	96.0%	estimated ± 2.4 pp, medium confidence
18	GLM-4.7	95.9%	measured
19	Kimi K2.6	95.9%	measured
20	Kimi K2.5	95.9%	measured
21	Kimi K2.5 (Reasoning)	95.9%	measured
22	Qwen 3.6 Max (preview)	95.9%	measured
23	Qwen3.8 Max Preview	95.6%	estimated ± 7.3 pp, medium confidence
24	Gemini 3.1 Pro	95.6%	measured
25	Step 5 Preview	95.6%	estimated ± 2.4 pp, medium confidence
26	GPT-5.6 Luna	95.4%	estimated ± 7.3 pp, medium confidence
27	Gemini 3.5 Flash	95.3%	measured
28	Qwen3.6-35B-A3B	95.3%	measured
29	Apodex 1.1	95.2%	estimated ± 7.3 pp, medium confidence
30	Apodex 1.1 Mini	95.2%	estimated ± 7.3 pp, medium confidence
31	MiMo-V2-Pro	95.0%	measured
32	Atria Dawn Preview	94.8%	estimated ± 14.1 pp, low confidence
33	Kimi K3	94.8%	estimated ± 2.4 pp, medium confidence
34	Qwen3.7 Max	94.7%	measured
35	Claude Opus 4.8	94.4%	measured
36	Gemini 3.7 Flash	94.3%	estimated ± 7.1 pp, low confidence
37	Qwen3.8-27B	94.3%	estimated ± 2.4 pp, medium confidence
38	Claude Sonnet 5.5	94.3%	estimated ± 7.1 pp, low confidence
39	Muse Spark 1.2	94.2%	estimated ± 13.3 pp, low confidence
40	MiMo-V2.5-Pro	94.2%	measured
41	Mistral Medium 3.5 128B	94.2%	measured
42	Qwen3.6-27B	94.2%	measured
43	Claude Opus 5	94.2%	estimated ± 7.1 pp, low confidence
44	Grok 4.5	94.1%	estimated ± 13.3 pp, low confidence
45	DeepSeek V4 Flash 0731	94.1%	estimated ± 13.3 pp, low confidence
46	GPT-6.1 Sol	94.1%	estimated ± 7.1 pp, low confidence
47	Qwen3.5-27B	93.9%	measured
48	Qwen3.5-122B-A10B	93.6%	measured
49	GPT-5.4 mini	93.4%	measured
50	Quasar 438B	93.3%	estimated ± 13.3 pp, low confidence
51	Grok 4.1 Fast (Reasoning)	93.3%	measured
52	Gemini 3.6 Flash	93.0%	estimated ± 13.3 pp, low confidence
53	Qwen3.7 Plus	93.0%	measured
54	Muse Spark 1.1	92.6%	estimated ± 13.3 pp, low confidence
55	GPT-5.4 nano	92.5%	measured
56	BTL-4	92.5%	estimated ± 14.1 pp, low confidence
57	Hy3	92.3%	estimated ± 13.3 pp, low confidence
58	Hy3 Preview	92.3%	estimated ± 13.3 pp, low confidence
59	Inkling-Small	92.2%	estimated ± 13.3 pp, low confidence
60	Claude Opus 4.6 (Adaptive)	92.1%	measured
61	GPT-5.2-Codex	92.1%	measured
62	Muse Spark	91.5%	measured
63	Ling 3.0 Flash	91.3%	estimated ± 13.3 pp, low confidence
64	Ling 3.0 Flash FP8	91.3%	estimated ± 13.3 pp, low confidence
65	MiMo-V2-Omni	91.2%	measured
66	Inkling	91.1%	estimated ± 2.4 pp, medium confidence
67	Pokee-Isaac 28B	90.7%	estimated ± 14.1 pp, low confidence
68	GLM-5.3	90.1%	estimated ± 2.4 pp, medium confidence
69	Kimi K2.7 Code	90.1%	measured
70	Trinity-Large-Preview	90.1%	measured
71	Trinity-Large-Thinking	90.1%	measured
72	Gemini 3.5 Flash-Lite	89.6%	estimated ± 13.3 pp, low confidence
73	Claude Opus 4.5 Thinking	89.5%	measured
74	Qwen3.5-35B-A3B	89.2%	measured
75	Muse Glimmer 30B	89.0%	estimated ± 2.4 pp, medium confidence
76	DeepSeek V4.1 Flash	89.0%	estimated ± 4.8 pp, medium confidence
77	MiniMax M3	88.9%	measured
78	Claude Opus 4.7 (Adaptive)	88.6%	measured
79	Claude Sonnet 5	88.6%	estimated ± 8.8 pp, low confidence
80	LFM2.5-8B-A1B	88.1%	measured
81	GLM-5.3-Flash	88.0%	estimated ± 2.4 pp, medium confidence
82	MiniCPM5-2B	87.6%	estimated ± 14.1 pp, low confidence
83	Gemini 3 Pro	87.1%	measured
84	GPT-5 (medium)	86.5%	measured
85	Claude Opus 4.5	86.3%	measured
86	GPT-5.6 Terra	86.3%	measured
87	Solar Pro 3	86.3%	measured
88	GPT-5.3 Codex	86.0%	measured
89	Ling 2.6 Flash	86.0%	measured
90	GPT-6 Astra	85.6%	estimated ± 4.8 pp, medium confidence
91	GPT-6 Sol	85.4%	estimated ± 4.8 pp, medium confidence
92	Claude Fable 5.1	85.4%	estimated ± 4.8 pp, medium confidence
93	Gemini 3.8 Flash	85.4%	estimated ± 4.8 pp, medium confidence
94	GPT-5.6 Sol	85.1%	measured
95	Command A+	85.0%	measured
96	Claude Opus 4.6	84.8%	measured
97	GPT-5.2	84.8%	measured
98	GPT-5 (high)	84.8%	measured
99	MiniMax M2.7	84.8%	measured
100	MiMo-V2-Flash	83.9%	measured
101	Qwen3.5 397B	83.9%	measured
102	Qwen3.5 397B (Reasoning)	83.9%	measured
103	Granite 4.2 30B	83.6%	estimated ± 14.1 pp, low confidence
104	Nemotron 3 Ultra	83.3%	measured
105	GPT-5.1-Codex	83.0%	measured
106	GPT-5.1-Codex-Max	83.0%	measured
107	GPT-5.1	81.9%	measured
108	o3	80.7%	measured
109	LLaDA2.2-flash	80.3%	measured
110	Ternary Bonsai 2 27B	80.2%	measured
111	LFM2.5-2.6B	79.9%	estimated ± 14.1 pp, low confidence
112	Nemotron 3.5 Lightning 30B A3B NVFP4	79.9%	estimated ± 13.3 pp, low confidence
113	Claude Sonnet 4.6	79.5%	measured
114	DeepSeek V3.2	78.9%	measured
115	Agents-A1-4B	78.2%	measured
116	GLM-4.6	76.9%	measured
117	Granite 4.2 3B	76.0%	estimated ± 14.1 pp, low confidence
118	Grok Code Fast 1	75.7%	measured
119	Grok 4	74.9%	measured
120	K-Exaone	74.3%	measured
121	Qwen3 Max	74.3%	measured
122	Claude Opus 4.7	74.0%	measured
123	Granite 4.2 8B	71.6%	estimated ± 13.3 pp, low confidence
124	Claude 4.1 Opus Thinking	71.4%	measured
125	Mellum2-12B-A2.5B-Thinking	69.6%	estimated ± 14.1 pp, low confidence
126	Mellum2-12B-A2.5B-Instruct	68.2%	estimated ± 14.1 pp, low confidence
127	Nemotron 3 Super 100B	67.8%	measured
128	GPT-OSS 120B	65.8%	measured
129	Grok 4 Fast (Reasoning)	65.8%	measured
130	Grok 4.1 Fast	63.7%	measured
131	ZAYA1-8B	62.9%	estimated ± 14.1 pp, low confidence
132	o1	62.6%	measured
133	Kimi K2	61.1%	measured
134	GPT-OSS 20B	60.2%	measured
135	Gemma 4 31B	59.9%	measured
136	LLaDA2.2-mini	57.5%	measured
137	LFM2.5-VL-3B	55.1%	estimated ± 14.1 pp, low confidence
138	Gemini 2.5 Pro	54.1%	measured
139	GPT-4.1 mini	52.9%	measured
140	Claude 4 Sonnet	52.3%	measured
141	DeepSeek V3 0324	47.1%	measured
142	GPT-4.1	47.1%	measured
143	Sarvam 105B	46.8%	measured
144	GLM-4.5-Air	46.5%	measured
145	MiniCPM5-1B	45.5%	estimated ± 14.1 pp, low confidence
146	Nemotron 3 Nano Omni 30B A3B	45.3%	measured
147	Gemma 4 26B A4B	43.6%	measured
148	Gemini 3 Flash	43.3%	measured
149	Mistral Small 4	41.2%	measured
150	Mistral Small 4 (Reasoning)	41.2%	measured
151	Nemotron 3 Nano 30B	40.9%	measured
152	LFM2.5-VL-450M	39.6%	estimated ± 14.1 pp, low confidence
153	LFM2.5-230M	39.5%	estimated ± 14.1 pp, low confidence
154	DeepSeek V3.1 (Reasoning)	37.4%	measured
155	North Mini Code	37.4%	measured
156	DeepSeek-R1	36.5%	measured
157	Gemma 4 12B	36.3%	measured
158	DeepSeek V3.1	34.8%	measured
159	Sarvam 30B	34.5%	measured
160	Celeris-1	32.0%	estimated ± 13.3 pp, low confidence
161	Solar Pro 2	31.9%	measured
162	Mistral Large 2	30.7%	measured
163	o3-mini	28.7%	measured
164	Ultravox v0.6 Llama 3.3 70B	26.6%	measured
165	GPT-4o	25.1%	measured
166	Mistral Large 3	24.6%	measured
167	Mistral Medium 3	24.3%	measured
168	DeepSeek V3	22.8%	measured
169	Qwen3-Omni-30B-A3B-Thinking	21.3%	measured
170	Claude 3 Haiku	21.1%	measured
171	Gemma 4 E2B	20.8%	measured
172	Gemma 4 E4B	20.8%	measured
173	Exaone 4.0 1.2B	20.5%	measured
174	Granite-4.0-H-1B	19.6%	measured
175	Llama 3.1 405B	19.0%	measured
176	Llama 4 Maverick	17.8%	measured
177	GPT-4.1 nano	17.3%	measured
178	Qwen3-Omni-30B-A3B-Instruct	16.4%	measured
179	Llama 4 Scout	15.5%	measured
180	Gemini 2.5 Flash	14.9%	measured
181	Granite-4.0-H-350M	14.6%	measured
182	Nova Pro	14.0%	measured
183	Granite-4.0-350M	13.2%	measured
184	Nemotron Ultra 253B	11.4%	measured
185	Gemma 3 27B	10.5%	measured
186	LFM2.5-VL-1.6B-Extract	8.5%	measured
187	Exaone 4.0 32B	4.1%	measured
188	Phi-4	0.0%	measured

τ²-bench results leaderboard

All leaderboards

Agentic · terminal

Agentic · tools

Coding

Math

Knowledge & reasoning

Instruction following

Multilingual

Vision & documents

Long context

General