benchgap
Agentic · tools

τ²-bench results leaderboard

As of 2026-11-22, the highest measured score on τ²-bench results is 99.1% by GLM-5.2. 75 more models have estimated scores, calibrated from the benchmarks they were measured on.

Measured scores: benchlm.ai.

#ModelScoreSource
1BTL-3100.0%estimated ± 12.4 pp, low confidence
2GLM-5.299.1%measured
3Kimi K3.1 Pro98.9%estimated ± 3.1 pp, medium confidence
4GPT-5.498.9%measured
5Claude Opus 5.598.9%estimated ± 3.1 pp, medium confidence
6Claude Opus 598.7%estimated ± 3.1 pp, medium confidence
7Claude Sonnet 5.598.7%estimated ± 3.1 pp, medium confidence
8DeepSeek V4.1 Flash 101798.7%estimated ± 3.1 pp, medium confidence
9LongCat-2.598.6%estimated ± 3.1 pp, medium confidence
10GLM-5.398.6%estimated ± 3.1 pp, medium confidence
11Claude Fable 598.5%measured
12GLM-5-Turbo98.5%measured
13GLM-5V-Turbo98.5%measured
14Step 3.7 Flash98.5%measured
15Muse Spark Open98.5%estimated ± 3.1 pp, medium confidence
16Hy498.3%estimated ± 3.1 pp, high confidence
17GLM-598.2%measured
18Claude Haiku 5.598.1%estimated ± 3.1 pp, high confidence
19Claude Sonnet 598.1%estimated ± 3.1 pp, high confidence
20GPT-6 Astra98.1%estimated ± 3.1 pp, high confidence
21Ling 3.0 Flash Sante98.1%estimated ± 3.1 pp, high confidence
22GPT-5.598.0%measured
23Qwen3.8 Max98.0%estimated ± 3.1 pp, high confidence
24Apodex 1.198.0%estimated ± 3.1 pp, high confidence
25Ornith-1.5-397B98.0%estimated ± 3.1 pp, high confidence
26Hy4 preview97.9%estimated ± 3.1 pp, high confidence
27GLM-5.3-Flash97.9%estimated ± 3.1 pp, high confidence
28Hy4 flash97.8%estimated ± 3.1 pp, high confidence
29Sage-A1.597.8%estimated ± 3.1 pp, high confidence
30GLM-5.197.7%measured
31Grok 4.397.7%measured
32Qwen3.6 Plus97.7%measured
33dots3-note Preview97.4%estimated ± 3.1 pp, high confidence
34Beam 1.0 501B-A23B97.0%estimated ± 3.1 pp, high confidence
35Beam 1.0 501B-A23B FP897.0%estimated ± 3.1 pp, high confidence
36Beam 1.0 501B-A23B NVFP497.0%estimated ± 3.1 pp, high confidence
37DeepSeek V4 Pro96.2%measured
38GLM-4.795.9%measured
39Kimi K2.695.9%measured
40Kimi K2.595.9%measured
41Qwen 3.6 Max (preview)95.9%measured
42Agents-A195.8%estimated ± 3.1 pp, high confidence
43Qwen4 Max95.7%estimated ± 7.0 pp, low confidence
44Gemini 3.1 Pro95.6%measured
45Qwen3.5 397B95.6%measured
46Qwen3.8 Max Preview95.5%estimated ± 7.0 pp, medium confidence
47Step 5 Preview95.4%estimated ± 7.0 pp, medium confidence
48Gemini 3.5 Flash95.3%measured
49Qwen3.6-35B-A3B95.3%measured
50Apodex 1.1 Mini95.1%estimated ± 7.0 pp, medium confidence
51MiMo-V2-Pro95.0%measured
52Ling 3.0 Flash Fin94.9%estimated ± 7.0 pp, medium confidence
53Qwen3.7 Max94.7%measured
54DeepSeek V4 Flash 073194.6%estimated ± 3.1 pp, high confidence
55Claude Opus 4.894.4%measured
56MiMo-V2.5-Pro94.2%measured
57Mistral Medium 3.5 128B94.2%measured
58Qwen3.6-27B94.2%measured
59Qwen3.5-27B93.9%measured
60Atria Dawn Preview93.7%estimated ± 11.8 pp, low confidence
61Ling 3.1 Flash93.6%estimated ± 11.8 pp, low confidence
62Qwen3.5-122B-A10B93.6%measured
63Inkling 2 Small93.6%estimated ± 11.8 pp, low confidence
64GPT-5.5 Pro93.6%estimated ± 11.8 pp, low confidence
65GPT-5.4 Pro93.6%estimated ± 11.8 pp, low confidence
66Claude Mythos 593.5%estimated ± 11.8 pp, low confidence
67ERNIE 5.5 Flash Preview93.5%estimated ± 11.8 pp, low confidence
68GPT-5.4 mini93.4%measured
69Mistral Large 493.4%estimated ± 11.8 pp, low confidence
70Grok 4.1 Fast (Reasoning)93.3%measured
71Gemini 3.8 Flash93.2%estimated ± 6.5 pp, low confidence
72Qwen3.7 Plus93.0%measured
73Inkling 2 Mini93.0%estimated ± 11.8 pp, low confidence
74Nemotron 3.5 Super VL 120B A12B92.9%estimated ± 11.8 pp, low confidence
75Nemotron 3.5 Super VL 120B A12B NVFP492.9%estimated ± 11.8 pp, low confidence
76Inkling-Small92.9%estimated ± 11.8 pp, low confidence
77BTL-492.9%estimated ± 12.4 pp, low confidence
78Inkling92.8%estimated ± 11.8 pp, low confidence
79Muse Spark 1.392.8%estimated ± 6.5 pp, low confidence
80Kimi K392.7%estimated ± 6.5 pp, low confidence
81GPT-5.4 nano92.5%measured
82Gemini 3.7 Flash92.4%estimated ± 6.5 pp, low confidence
83Ling 3.0 Flash92.2%estimated ± 11.8 pp, low confidence
84Claude Opus 4.6 (Adaptive)92.1%measured
85GPT-5.2-Codex92.1%measured
86Qwen3.5-4B92.1%measured
87Grok 4.691.9%estimated ± 6.5 pp, low confidence
88Nemotron 3.5 Bolt 74B-A9B BF1691.9%estimated ± 11.8 pp, low confidence
89Muse Spark91.5%measured
90MiMo-V2-Omni91.2%measured
91Nemotron 3.5 Bolt 74B-A9B91.2%estimated ± 11.8 pp, low confidence
92Pokee-Isaac 28B91.1%estimated ± 12.4 pp, low confidence
93Kimi K2.7 Code90.1%measured
94Trinity-Large-Preview90.1%measured
95Trinity-Large-Thinking90.1%measured
96GPT-5.6 Luna89.6%estimated ± 6.5 pp, low confidence
97Claude Opus 4.5 Thinking89.5%measured
98Qwen3.5-35B-A3B89.2%measured
99MiniMax M388.9%measured
100Claude Opus 4.788.6%measured
101LFM2.5-8B-A1B88.1%measured
102MiniCPM5-2B88.0%estimated ± 12.4 pp, low confidence
103Gemini 3 Pro87.1%measured
104Spark-X2.6-4B86.9%measured
105LFM2.5-24B-A2B86.8%estimated ± 12.4 pp, low confidence
106Claude Opus 4.586.3%measured
107GPT-5.6 Terra86.3%measured
108Solar Pro 386.3%measured
109GPT-5.3 Codex86.0%measured
110Ling 2.6 Flash86.0%measured
111GPT-5.6 Sol85.1%measured
112Command A+85.0%measured
113Claude Opus 4.684.8%measured
114GPT-584.8%measured
115GPT-5.284.8%measured
116MiniMax M2.784.8%measured
117Granite 4.2 30B84.0%estimated ± 12.4 pp, low confidence
118MiMo-V2-Flash83.9%measured
119Nemotron 3 Ultra83.3%measured
120GPT-5.1-Codex83.0%measured
121GPT-5.1-Codex-Max83.0%measured
122GPT-5.181.9%measured
123o380.7%measured
124LLaDA2.2-flash80.3%measured
125LFM2.5-2.6B80.3%estimated ± 12.4 pp, low confidence
126Ternary Bonsai 2 27B80.2%measured
127Claude Sonnet 4.679.5%measured
128DeepSeek V3.278.9%measured
129Solar Pro 478.5%estimated ± 11.8 pp, low confidence
130Agents-A1-4B78.2%measured
131LongCat-Flash-Lite-Sparse77.5%estimated ± 11.8 pp, low confidence
132GLM-4.676.9%measured
133Granite 4.2 3B76.4%estimated ± 12.4 pp, low confidence
134Granite 4.2 8B76.4%estimated ± 12.4 pp, low confidence
135Ornith-1.5-35B-A3B76.2%estimated ± 3.1 pp, medium confidence
136Grok Code Fast 175.7%measured
137Grok 474.9%measured
138K-Exaone74.3%measured
139Qwen3 Max74.3%measured
140Claude 4 Opus Thinking73.4%measured
141Mellum2-12B-A2.5B-Thinking69.9%estimated ± 12.4 pp, low confidence
142Mellum2-12B-A2.5B-Instruct68.5%estimated ± 12.4 pp, low confidence
143GPT-OSS 120B65.8%measured
144Grok 4 Fast (Reasoning)65.8%measured
145Ornith-1.5-9B65.2%estimated ± 3.1 pp, medium confidence
146Grok 4.1 Fast63.7%measured
147ZAYA1-8B63.2%estimated ± 12.4 pp, low confidence
148o162.6%measured
149Kimi K261.1%measured
150GPT-OSS 20B60.2%measured
151Gemma 4 31B59.9%measured
152LLaDA2.2-mini57.5%measured
153o4-mini55.6%measured
154LFM2.5-VL-3B55.4%estimated ± 12.4 pp, low confidence
155Nemotron 3.5 Bolt E4B BF1655.1%measured
156Claude 3.7 Sonnet Thinking54.7%measured
157Gemini 2.5 Pro54.1%measured
158GPT-4.1 mini52.9%measured
159Nemotron 3.5 Bolt E4B52.3%measured
160Claude 4 Sonnet52.3%measured
161Claude 3.7 Sonnet50.0%measured
162DeepSeek V3 032447.1%measured
163GPT-4.147.1%measured
164Sarvam 105B46.8%measured
165GLM-4.5-Air46.5%measured
166MiniCPM5-1B45.7%estimated ± 12.4 pp, low confidence
167Nemotron 3 Nano Omni 30B A3B45.3%measured
168Nemotron 3.5 Lightning 30B A3B NVFP444.1%estimated ± 11.8 pp, low confidence
169Gemma 4 26B A4B43.6%measured
170Gemini 3 Flash43.3%measured
171Mistral Small 441.2%measured
172LFM2.5-VL-450M39.8%estimated ± 12.4 pp, low confidence
173LFM2.5-230M39.7%estimated ± 12.4 pp, low confidence
174North Mini Code37.4%measured
175DeepSeek-R136.5%measured
176Gemma 4 12B36.3%measured
177DeepSeek V3.134.8%measured
178Sarvam 30B34.5%measured
179Solar Pro 231.9%measured
180Mistral Large 230.7%measured
181Gemini 2.0 Flash29.5%measured
182o3-mini28.7%measured
183Devstral Small28.4%measured
184Llama 3.3 70B26.6%measured
185Magistral Small26.6%measured
186Ultravox v0.6 Llama 3.3 70B26.6%measured
187Nemotron 3 Nano 30B25.4%measured
188GPT-4o25.1%measured
189Mistral Small 3.125.1%measured
190Claude 3.5 Haiku24.6%measured
191Mistral Large 324.6%measured
192Mistral Medium 324.3%measured
193Magistral Medium23.1%measured
194DeepSeek V322.8%measured
195Llama 3.1 70B21.6%measured
196Qwen3-Omni-30B-A3B-Thinking21.3%measured
197Claude 3 Haiku21.1%measured
198Gemma 4 E2B20.8%measured
199Gemma 4 E4B20.8%measured
200Exaone 4.0 1.2B20.5%measured
201Granite-4.0-H-1B19.6%measured
202Gemini 2.5 Flash-Lite19.0%measured
203Llama 3.1 405B19.0%measured
204Llama 4 Maverick17.8%measured
205GPT-4.1 nano17.3%measured
206Qwen3-Omni-30B-A3B-Instruct16.4%measured
207Llama 4 Scout15.5%measured
208Gemini 2.5 Flash14.9%measured
209Granite-4.0-H-350M14.6%measured
210Nova Pro14.0%measured
211Granite-4.0-350M13.2%measured
212Nemotron Ultra 253B11.4%measured
213Gemma 3 27B10.5%measured
214LFM2.5-VL-1.6B-Extract8.5%measured
215Exaone 4.0 32B4.1%measured
216LFM3-1.7B1.7%estimated ± 3.1 pp, medium confidence
217Phi-40.0%measured
218Phi-4-reasoning0.0%measured

All leaderboards

Agentic · terminal

Agentic · tools

Coding

Math

Knowledge & reasoning

Instruction following

Multilingual

Vision & documents

Long context

General