Skip to main content
BenchLM
Data

τ²-Bench Tool-Agent-User Evaluation (τ²-bench results)

We show this table for reference; we do not rank on it.

Data verified

This route is a sourced ledger for published τ²-bench results. Most current rows come from Artificial Analysis's telecom implementation, while named provider rows can use telecom, airline, retail, or aggregate setups.

Benchmark score on τ²-bench sourced results — November 12, 2026

We compile the τ²-bench sourced results rows from provider self-reports and secondary reports. The table contains 168 models. GLM-5.2 has the highest published score at 99.1%, but the available coverage and evaluation setups do not establish a market leader. We do not use these results to rank models overall.

1Open weights

GLM-5.2

Z.AI

Carried from an upstream page

Artificial Analysis τ²-bench

99.1%
Overall 59.88Context 1M
2Closed

GPT-5.4

OpenAI

Reported by OpenAI

τ²-bench Telecom

98.9%
Overall 65.62Context 1.05M
3Open weights

GLM-4.7-Flash

Z.AI

Carried from an upstream page

Artificial Analysis τ²-bench

98.8%
Overall 38.87Context 200K

168 modelsAgenticCurrentDisplay onlyUpdated November 12, 2026

Benchmark score table (168 models)

None of the 168 results on this page was run by the benchmark or an independent evaluator: 15 reported by the labs that built the models, 4 reported by a third party and 149 carried from an upstream page, where who ran them is not recorded.

Benchmark score results for τ²-bench results
RankModel / configurationScoreParameters (B)Open / closed
1GLM-5.2Z.AICarried from an upstream pageArtificial Analysis τ²-bench
99.1%
Not reportedOpen
2GPT-5.4OpenAIReported by OpenAIτ²-bench Telecom
98.9%
Not reportedClosed
3GLM-4.7-FlashZ.AICarried from an upstream pageArtificial Analysis τ²-bench
98.8%
Not reportedOpen
4Claude Fable 5AnthropicCarried from an upstream pageArtificial Analysis τ²-bench
98.5%
Not reportedClosed
5GLM-5-TurboZ.AICarried from an upstream pageArtificial Analysis τ²-bench
98.5%
Not reportedClosed
6GLM-5V-TurboZ.AICarried from an upstream pageArtificial Analysis τ²-bench
98.5%
Not reportedClosed
7Step 3.7 FlashStepFunCarried from an upstream pageArtificial Analysis τ²-bench
98.5%
Not reportedOpen
8GLM-5Z.AICarried from an upstream pageArtificial Analysis τ²-bench
98.2%
Not reportedOpen
9GPT-5.5OpenAIReported by OpenAIτ²-bench Telecom
98%
Not reportedClosed
10Qwen3.6 PlusAlibabaCarried from an upstream pageArtificial Analysis τ²-bench
97.7%
Not reportedClosed
11GLM-5.1Z.AICarried from an upstream pageArtificial Analysis τ²-bench
97.7%
Not reportedOpen
12Grok 4.3xAIReported by a third partyArtificial Analysis τ²-bench
97.7%
Not reportedClosed
13DeepSeek V4 Pro 0813DeepSeekCarried from an upstream pageArtificial Analysis τ²-bench
96.2%
Not reportedOpen
14Kimi K2.6Moonshot AICarried from an upstream pageArtificial Analysis τ²-bench
95.9%
Not reportedOpen
15Kimi K2.5 (Reasoning)Moonshot AICarried from an upstream pageArtificial Analysis τ²-bench
95.9%
Not reportedOpen
16Kimi K2.5Moonshot AICarried from an upstream pageArtificial Analysis τ²-bench
95.9%
Not reportedOpen
17GLM-4.7Z.AICarried from an upstream pageArtificial Analysis τ²-bench
95.9%
Not reportedOpen
18Qwen 3.6 Max (preview)AlibabaCarried from an upstream pageArtificial Analysis τ²-bench
95.9%
Not reportedClosed
19Gemini 3.1 ProGoogleReported by Meta AI (comparison table)τ²-bench published setup
95.6%
Not reportedClosed
20Qwen3.5 397BAlibabaCarried from an upstream pageArtificial Analysis τ²-bench
95.6%
Not reportedOpen
21Qwen3.5 397B (Reasoning)AlibabaCarried from an upstream pageArtificial Analysis τ²-bench
95.6%
Not reportedOpen
22Gemini 3.5 FlashGoogleReported by a third partyArtificial Analysis τ²-bench
95.3%
Not reportedClosed
23Qwen3.6-35B-A3BAlibabaCarried from an upstream pageArtificial Analysis τ²-bench
95.3%
Not reportedOpen
24MiniMax M2.5MiniMaxCarried from an upstream pageArtificial Analysis τ²-bench
95.3%
Not reportedClosed
25MiMo-V2-ProXiaomiCarried from an upstream pageArtificial Analysis τ²-bench
95%
Not reportedClosed
26Qwen3.7 MaxAlibabaCarried from an upstream pageArtificial Analysis τ²-bench
94.7%
Not reportedClosed
27Claude Opus 4.8AnthropicCarried from an upstream pageArtificial Analysis τ²-bench
94.4%
Not reportedClosed
28Qwen3.6-27BAlibabaCarried from an upstream pageArtificial Analysis τ²-bench
94.2%
Not reportedOpen
29MiMo-V2.5-ProXiaomiCarried from an upstream pageArtificial Analysis τ²-bench
94.2%
Not reportedClosed
30Mistral Medium 3.5 128BMistralCarried from an upstream pageArtificial Analysis τ²-bench
94.2%
Not reportedOpen
31DeepSeek V4 Pro (High)DeepSeekCarried from an upstream pageArtificial Analysis τ²-bench
94.2%
Not reportedOpen
32Qwen3.5-27BAlibabaCarried from an upstream pageArtificial Analysis τ²-bench
93.9%
Not reportedOpen
33Qwen3.5-122B-A10BAlibabaCarried from an upstream pageArtificial Analysis τ²-bench
93.6%
Not reportedOpen
34GPT-5.4 miniOpenAIReported by OpenAIτ²-bench Telecom
93.4%
Not reportedClosed
35Grok 4.1 Fast (Reasoning)xAICarried from an upstream pageArtificial Analysis τ²-bench
93.3%
Not reportedClosed
36Qwen3.7 PlusAlibabaCarried from an upstream pageArtificial Analysis τ²-bench
93%
Not reportedClosed
37GPT-5.4 nanoOpenAIReported by OpenAIτ²-bench Telecom
92.5%
Not reportedClosed
38GPT-5.2-CodexOpenAICarried from an upstream pageArtificial Analysis τ²-bench
92.1%
Not reportedClosed
39Claude Opus 4.6 (Adaptive)AnthropicCarried from an upstream pageArtificial Analysis τ²-bench
92.1%
Not reportedClosed
40Qwen3.5-4BAlibabaCarried from an upstream pageArtificial Analysis τ²-bench
92.1%
Not reportedOpen
41Muse SparkMetaReported by Metaτ²-bench published setup
91.5%
Not reportedClosed
42MiMo-V2-OmniXiaomiCarried from an upstream pageArtificial Analysis τ²-bench
91.2%
Not reportedClosed
43Kimi K2.7 CodeMoonshot AICarried from an upstream pageArtificial Analysis τ²-bench
90.1%
Not reportedOpen
44Trinity-Large-ThinkingArcee AICarried from an upstream pageArtificial Analysis τ²-bench
90.1%
Not reportedOpen
45Trinity-Large-PreviewArcee AICarried from an upstream pageArtificial Analysis τ²-bench
90.1%
Not reportedOpen
46Claude Opus 4.5 ThinkingAnthropicCarried from an upstream pageArtificial Analysis τ²-bench
89.5%
Not reportedClosed
47Qwen3.5-35B-A3BAlibabaCarried from an upstream pageArtificial Analysis τ²-bench
89.2%
Not reportedOpen
48MiniMax M3MiniMaxCarried from an upstream pageArtificial Analysis τ²-bench
88.9%
Not reportedOpen
49Claude Opus 4.7 (Adaptive)AnthropicCarried from an upstream pageArtificial Analysis τ²-bench
88.6%
Not reportedClosed
50LFM2.5-8B-A1BLiquidAIReported by LiquidAIτ²-bench Telecom
88.1%
Not reportedOpen
51Step 3.5 FlashStepFunCarried from an upstream pageArtificial Analysis τ²-bench
87.4%
Not reportedOpen
52Gemini 3 ProGoogleCarried from an upstream pageArtificial Analysis τ²-bench
87.1%
Not reportedClosed
53Spark-X2.6-4BSparkLLMReported by SparkLLMτ²-bench published setup
86.9%
Not reportedOpen
54GPT-5 (medium)OpenAICarried from an upstream pageArtificial Analysis τ²-bench
86.5%
Not reportedClosed
55GPT-5.6 TerraOpenAICarried from an upstream pageArtificial Analysis τ²-bench
86.3%
Not reportedClosed
56Claude Opus 4.5AnthropicCarried from an upstream pageArtificial Analysis τ²-bench
86.3%
Not reportedClosed
57Solar Pro 3UpstageCarried from an upstream pageArtificial Analysis τ²-bench
86.3%
Not reportedClosed
58GPT-5.3 CodexOpenAICarried from an upstream pageArtificial Analysis τ²-bench
86%
Not reportedClosed
59Ling 2.6 FlashInclusionAIReported by a third partyArtificial Analysis τ²-bench
86%
Not reportedOpen
60GPT-5.3-Codex-SparkOpenAICarried from an upstream pageArtificial Analysis τ²-bench
86%
Not reportedClosed
61GPT-5.6 SolOpenAICarried from an upstream pageArtificial Analysis τ²-bench
85.1%
Not reportedClosed
62Command A+CohereReported by Cohereτ²-bench Telecom
85%
Not reportedOpen
63Claude Opus 4.6AnthropicCarried from an upstream pageArtificial Analysis τ²-bench
84.8%
Not reportedClosed
64GPT-5.2OpenAICarried from an upstream pageArtificial Analysis τ²-bench
84.8%
Not reportedClosed
65MiniMax M2.7MiniMaxCarried from an upstream pageArtificial Analysis τ²-bench
84.8%
Not reportedOpen
66GPT-5 (high)OpenAICarried from an upstream pageArtificial Analysis τ²-bench
84.8%
Not reportedClosed
67MiMo-V2-FlashXiaomiCarried from an upstream pageArtificial Analysis τ²-bench
83.9%
Not reportedOpen
68Nemotron 3 UltraNVIDIACarried from an upstream pageArtificial Analysis τ²-bench
83.3%
Not reportedOpen
69GPT-5.1-CodexOpenAICarried from an upstream pageArtificial Analysis τ²-bench
83%
Not reportedClosed
70GPT-5.1-Codex-MaxOpenAICarried from an upstream pageArtificial Analysis τ²-bench
83%
Not reportedClosed
71GPT-5.1OpenAICarried from an upstream pageArtificial Analysis τ²-bench
81.9%
Not reportedClosed
72o3OpenAICarried from an upstream pageArtificial Analysis τ²-bench
80.7%
Not reportedClosed
73LLaDA2.2-flashInclusionAIReported by InclusionAIτ²-bench published setup
80.3%
Not reportedOpen
74Ternary Bonsai 2 27BPrism MLReported by Prism MLτ²-bench Telecom
80.2%
Not reportedOpen
75Claude Sonnet 4.6AnthropicCarried from an upstream pageArtificial Analysis τ²-bench
79.5%
Not reportedClosed
76DeepSeek V3.2DeepSeekCarried from an upstream pageArtificial Analysis τ²-bench
78.9%
Not reportedOpen
77Agents-A1-4BInternScienceReported by InternScienceτ²-bench published setup
78.2%
Not reportedOpen
78GLM-4.6Z.AICarried from an upstream pageArtificial Analysis τ²-bench
76.9%
Not reportedOpen
79Grok Code Fast 1xAICarried from an upstream pageArtificial Analysis τ²-bench
75.7%
Not reportedClosed
80Grok 4xAICarried from an upstream pageArtificial Analysis τ²-bench
74.9%
Not reportedClosed
81Qwen3 MaxAlibabaCarried from an upstream pageArtificial Analysis τ²-bench
74.3%
Not reportedClosed
82K-ExaoneLG AI ResearchCarried from an upstream pageArtificial Analysis τ²-bench
74.3%
Not reportedClosed
83Claude Opus 4.7AnthropicCarried from an upstream pageArtificial Analysis τ²-bench
74%
Not reportedClosed
84Claude 4 Opus ThinkingAnthropicCarried from an upstream pageArtificial Analysis τ²-bench
73.4%
Not reportedClosed
85Claude 4.1 Opus ThinkingAnthropicCarried from an upstream pageArtificial Analysis τ²-bench
71.4%
Not reportedClosed
86Mercury 2InceptionCarried from an upstream pageArtificial Analysis τ²-bench
70.8%
Not reportedClosed
87GPT-5 miniOpenAICarried from an upstream pageArtificial Analysis τ²-bench
68.4%
Not reportedClosed
88GPT-OSS 120BOpenAICarried from an upstream pageArtificial Analysis τ²-bench
65.8%
Not reportedOpen
89Grok 4 Fast (Reasoning)xAICarried from an upstream pageArtificial Analysis τ²-bench
65.8%
Not reportedClosed
90Grok 4.1 FastxAICarried from an upstream pageArtificial Analysis τ²-bench
63.7%
Not reportedClosed
91o1OpenAICarried from an upstream pageArtificial Analysis τ²-bench
62.6%
Not reportedClosed
92Kimi K2Moonshot AICarried from an upstream pageArtificial Analysis τ²-bench
61.1%
Not reportedClosed
93GPT-OSS 20BOpenAICarried from an upstream pageArtificial Analysis τ²-bench
60.2%
Not reportedOpen
94Gemma 4 31BGoogleCarried from an upstream pageArtificial Analysis τ²-bench
59.9%
Not reportedOpen
95LLaDA2.2-miniInclusionAIReported by InclusionAIτ²-bench published setup
57.5%
Not reportedOpen
96o4-miniOpenAICarried from an upstream pageArtificial Analysis τ²-bench
55.6%
Not reportedClosed
97Nemotron 3.5 Bolt E4B BF16NVIDIAReported by NVIDIAτ²-bench published setup
55.1%
Not reportedOpen
98Claude 3.7 Sonnet ThinkingAnthropicCarried from an upstream pageArtificial Analysis τ²-bench
54.7%
Not reportedClosed
99Gemini 2.5 ProGoogleCarried from an upstream pageArtificial Analysis τ²-bench
54.1%
Not reportedClosed
100GPT-4.1 miniOpenAICarried from an upstream pageArtificial Analysis τ²-bench
52.9%
Not reportedClosed
101Nemotron 3.5 Bolt E4BNVIDIAReported by NVIDIAτ²-bench published setup
52.3%
Not reportedOpen
102Claude 4 SonnetAnthropicCarried from an upstream pageArtificial Analysis τ²-bench
52.3%
Not reportedClosed
103Claude 3.7 SonnetAnthropicCarried from an upstream pageArtificial Analysis τ²-bench
50%
Not reportedClosed
104DeepSeek V3 0324DeepSeekCarried from an upstream pageArtificial Analysis τ²-bench
47.1%
Not reportedOpen
105GPT-4.1OpenAICarried from an upstream pageArtificial Analysis τ²-bench
47.1%
Not reportedClosed
106Sarvam 105BSarvamCarried from an upstream pageArtificial Analysis τ²-bench
46.8%
Not reportedOpen
107GLM-4.5-AirZ.AICarried from an upstream pageArtificial Analysis τ²-bench
46.5%
Not reportedClosed
108Nemotron 3 Nano Omni 30B A3BNVIDIAReported by NVIDIAτ²-bench Telecom
45.3%
Not reportedOpen
109Gemma 4 26B A4BGoogleCarried from an upstream pageArtificial Analysis τ²-bench
43.6%
Not reportedOpen
110Gemini 3 FlashGoogleCarried from an upstream pageArtificial Analysis τ²-bench
43.3%
Not reportedClosed
111Mistral Small 4MistralCarried from an upstream pageArtificial Analysis τ²-bench
41.2%
Not reportedOpen
112Mistral Small 4 (Reasoning)MistralCarried from an upstream pageArtificial Analysis τ²-bench
41.2%
Not reportedOpen
113DeepSeek V3.1 (Reasoning)DeepSeekCarried from an upstream pageArtificial Analysis τ²-bench
37.4%
Not reportedOpen
114North Mini CodeCohereCarried from an upstream pageArtificial Analysis τ²-bench
37.4%
Not reportedOpen
115DeepSeek-R1DeepSeekCarried from an upstream pageArtificial Analysis τ²-bench
36.5%
Not reportedOpen
116GPT-5 nanoOpenAICarried from an upstream pageArtificial Analysis τ²-bench
36.5%
Not reportedClosed
117Gemma 4 12BGoogleCarried from an upstream pageArtificial Analysis τ²-bench
36.3%
Not reportedOpen
118DeepSeek V3.1DeepSeekCarried from an upstream pageArtificial Analysis τ²-bench
34.8%
Not reportedOpen
119Sarvam 30BSarvamCarried from an upstream pageArtificial Analysis τ²-bench
34.5%
Not reportedOpen
120MiniMax M1 80kMiniMaxCarried from an upstream pageArtificial Analysis τ²-bench
34.2%
Not reportedClosed
121Solar Pro 2UpstageCarried from an upstream pageArtificial Analysis τ²-bench
31.9%
Not reportedClosed
122Mistral Large 2MistralCarried from an upstream pageArtificial Analysis τ²-bench
30.7%
Not reportedClosed
123Gemini 2.0 FlashGoogleCarried from an upstream pageArtificial Analysis τ²-bench
29.5%
Not reportedClosed
124o3-miniOpenAICarried from an upstream pageArtificial Analysis τ²-bench
28.7%
Not reportedClosed
125Devstral SmallMistralCarried from an upstream pageArtificial Analysis τ²-bench
28.4%
Not reportedOpen
126Ministral 3 14B (Reasoning)MistralCarried from an upstream pageArtificial Analysis τ²-bench
27.2%
Not reportedOpen
127Ministral 3 14BMistralCarried from an upstream pageArtificial Analysis τ²-bench
27.2%
Not reportedOpen
128Llama 3.3 70BMetaCarried from an upstream pageArtificial Analysis τ²-bench
26.6%
Not reportedOpen
129Magistral SmallMistralCarried from an upstream pageArtificial Analysis τ²-bench
26.6%
Not reportedOpen
130Ultravox v0.6 Llama 3.3 70BFixie AICarried from an upstream pageArtificial Analysis τ²-bench
26.6%
Not reportedOpen
131Ministral 3 8B (Reasoning)MistralCarried from an upstream pageArtificial Analysis τ²-bench
26.6%
Not reportedOpen
132Ministral 3 8BMistralCarried from an upstream pageArtificial Analysis τ²-bench
26.6%
Not reportedOpen
133Nemotron 3 Nano 30BNVIDIACarried from an upstream pageArtificial Analysis τ²-bench
25.4%
Not reportedOpen
134GPT-4oOpenAICarried from an upstream pageArtificial Analysis τ²-bench
25.1%
Not reportedClosed
135Mistral Small 3.1MistralCarried from an upstream pageArtificial Analysis τ²-bench
25.1%
Not reportedOpen
136Ministral 3 3B (Reasoning)MistralCarried from an upstream pageArtificial Analysis τ²-bench
24.9%
Not reportedOpen
137Ministral 3 3BMistralCarried from an upstream pageArtificial Analysis τ²-bench
24.9%
Not reportedOpen
138Mistral Large 3MistralCarried from an upstream pageArtificial Analysis τ²-bench
24.6%
Not reportedOpen
139Claude 3.5 HaikuAnthropicCarried from an upstream pageArtificial Analysis τ²-bench
24.6%
Not reportedClosed
140Mistral Medium 3MistralCarried from an upstream pageArtificial Analysis τ²-bench
24.3%
Not reportedClosed
141Magistral MediumMistralCarried from an upstream pageArtificial Analysis τ²-bench
23.1%
Not reportedClosed
142DeepSeek V3DeepSeekCarried from an upstream pageArtificial Analysis τ²-bench
22.8%
Not reportedOpen
143Llama 3.1 70BMetaCarried from an upstream pageArtificial Analysis τ²-bench
21.6%
Not reportedOpen
144Qwen3-Omni-30B-A3B-ThinkingAlibabaCarried from an upstream pageArtificial Analysis τ²-bench
21.3%
Not reportedOpen
145Claude 3 HaikuAnthropicCarried from an upstream pageArtificial Analysis τ²-bench
21.1%
Not reportedClosed
146Gemma 4 E4BGoogleCarried from an upstream pageArtificial Analysis τ²-bench
20.8%
Not reportedOpen
147Gemma 4 E2BGoogleCarried from an upstream pageArtificial Analysis τ²-bench
20.8%
Not reportedOpen
148Exaone 4.0 1.2BLG AI ResearchCarried from an upstream pageArtificial Analysis τ²-bench
20.5%
Not reportedOpen
149Granite-4.0-H-1BIBMCarried from an upstream pageArtificial Analysis τ²-bench
19.6%
Not reportedOpen
150LFM2.5-1.2B-ThinkingLiquidAICarried from an upstream pageArtificial Analysis τ²-bench
19.6%
Not reportedClosed
151Gemini 2.5 Flash-LiteGoogleCarried from an upstream pageArtificial Analysis τ²-bench
19%
Not reportedClosed
152Llama 3.1 405BMetaCarried from an upstream pageArtificial Analysis τ²-bench
19%
Not reportedOpen
153Llama 4 MaverickMetaCarried from an upstream pageArtificial Analysis τ²-bench
17.8%
Not reportedOpen
154GPT-4.1 nanoOpenAICarried from an upstream pageArtificial Analysis τ²-bench
17.3%
Not reportedClosed
155Qwen3-Omni-30B-A3B-InstructAlibabaCarried from an upstream pageArtificial Analysis τ²-bench
16.4%
Not reportedOpen
156Llama 4 ScoutMetaCarried from an upstream pageArtificial Analysis τ²-bench
15.5%
Not reportedOpen
157Gemini 2.5 FlashGoogleCarried from an upstream pageArtificial Analysis τ²-bench
14.9%
Not reportedClosed
158Granite-4.0-H-350MIBMCarried from an upstream pageArtificial Analysis τ²-bench
14.6%
Not reportedOpen
159Nova ProAmazonCarried from an upstream pageArtificial Analysis τ²-bench
14%
Not reportedClosed
160Granite-4.0-350MIBMCarried from an upstream pageArtificial Analysis τ²-bench
13.2%
Not reportedOpen
161Nemotron Ultra 253BNVIDIACarried from an upstream pageArtificial Analysis τ²-bench
11.4%
Not reportedOpen
162LFM2-24B-A2BLiquidAICarried from an upstream pageArtificial Analysis τ²-bench
11.1%
Not reportedClosed
163LFM2.5-1.2B-InstructLiquidAICarried from an upstream pageArtificial Analysis τ²-bench
10.8%
Not reportedClosed
164Gemma 3 27BGoogleCarried from an upstream pageArtificial Analysis τ²-bench
10.5%
Not reportedOpen
165LFM2.5-VL-1.6B-ExtractLiquidAICarried from an upstream pageArtificial Analysis τ²-bench
8.5%
Not reportedOpen
166Exaone 4.0 32BLG AI ResearchCarried from an upstream pageArtificial Analysis τ²-bench
4.1%
Not reportedOpen
167Phi-4MicrosoftCarried from an upstream pageArtificial Analysis τ²-bench
0%
Not reportedOpen
168Phi-4-reasoningMicrosoftCarried from an upstream pageArtificial Analysis τ²-bench
0%
Not reportedOpen
  • Lab-reported
  • Benchmark-run
  • Independent
What the labels mean

How to read this leaderboard

Editorial review by Glevd · 2026-07-15

Use a row only with its attached source and setup label. Match the domain, task release, agent model, user-simulator model, scaffold, prompts, trial count, and pass^k metric before comparing scores. The sorted table is a source ledger, not a controlled cross-provider ranking.

Operator receipt: 168 sourced rows are currently displayable on this page; the highest published score among these 168 models is GLM-5.2 at 99.1%, which does not establish a market leader.

Honest limit: The page mixes a large third-party telecom snapshot with smaller provider-published slices. Those sources do not use one guaranteed-common harness or reporting policy, and BenchLM did not rerun them. A higher number can reflect a different user model, prompt, domain, task release, or repeat policy.

Original TAU-bench methodτ²-bench Airline ledgerτ³-bench ledgerHow benchmark confidence works
Maintained τ-bench repositoryτ²-bench paperOfficial leaderboardArtificial Analysis evaluation

Among the reported τ²-bench results rows, GLM-5.2 is first at 99.1%. The third row is 0.3 points behind. The broader top-10 range is 1.4 points, so many of the published results sit in a relatively narrow band.

168 models have been evaluated on τ²-bench results. The benchmark falls in the Agentic category. τ²-bench results is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About τ²-bench results

Year

2025

Tasks

Airline, retail, and telecom customer-service task sets

Format

Published domain success or pass^k results

Difficulty

Dual-control customer-service workflows

τ²-bench extends the original benchmark with a dual-control telecom domain where the agent and simulated user can both act through tools. The maintained framework also includes airline and retail. BenchLM keeps each exact source attached and labels the published setup instead of treating every row as one controlled run.

τ²-Bench: Evaluating Conversational Agents in a Dual-Control Environment

Freshness and provenance

Version

τ²-Bench 2026

Refresh cadence

Quarterly

Staleness state

Current

Question availability

Public benchmark set

CurrentDisplay only

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

Questions

Are all τ²-bench scores directly comparable?

No. Match the domain, task release, agent model, user model, scaffold, prompts, number of trials, and pass^k definition. BenchLM labels each sourced row so a telecom result or third-party implementation is not silently treated as the same setup as an airline, retail, or aggregate result.

Does a high τ²-bench score prove production support reliability?

No. It measures success in simulated customer-service environments under a reported setup. Production identity checks, permission boundaries, changing policies, latency, cost, monitoring, and human escalation still need separate testing.

Compare top models on τ²-bench results

GLM-5.2 vs GPT-5.4GPT-5.4 vs GLM-4.7-FlashGLM-4.7-Flash vs Claude Fable 5Claude Fable 5 vs GLM-5-Turbo
Last updated: November 12, 2026 · BenchLM version τ²-Bench 2026

Know when it’s worth switching models

The model to choose, the cheaper alternative, and the release we would wait on.

Read a sample issue

Join 7,000+ subscribers.

One email each week. Unsubscribe anytime.