τ²-Bench Tool-Agent-User Evaluation (τ²-bench results)
We show this table for reference; we do not rank on it.
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 — October 28, 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.
GLM-5.2
Z.AI
Artificial Analysis τ²-bench
GPT-5.4
OpenAI
τ²-bench Telecom
GLM-4.7-Flash
Z.AI
Artificial Analysis τ²-bench
168 modelsAgenticCurrentDisplay onlyUpdated October 28, 2026
Benchmark score table (168 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | GLM-5.2Z.AIArtificial Analysis τ²-bench | 99.1% | Not reported | Open |
| 2 | GPT-5.4OpenAIτ²-bench Telecom | 98.9% | Not reported | Closed |
| 3 | GLM-4.7-FlashZ.AIArtificial Analysis τ²-bench | 98.8% | Not reported | Open |
| 4 | Claude Fable 5AnthropicArtificial Analysis τ²-bench | 98.5% | Not reported | Closed |
| 5 | GLM-5-TurboZ.AIArtificial Analysis τ²-bench | 98.5% | Not reported | Closed |
| 6 | GLM-5V-TurboZ.AIArtificial Analysis τ²-bench | 98.5% | Not reported | Closed |
| 7 | Step 3.7 FlashStepFunArtificial Analysis τ²-bench | 98.5% | Not reported | Open |
| 8 | GLM-5Z.AIArtificial Analysis τ²-bench | 98.2% | Not reported | Open |
| 9 | GPT-5.5OpenAIτ²-bench Telecom | 98% | Not reported | Closed |
| 10 | Qwen3.6 PlusAlibabaArtificial Analysis τ²-bench | 97.7% | Not reported | Closed |
| 11 | GLM-5.1Z.AIArtificial Analysis τ²-bench | 97.7% | Not reported | Open |
| 12 | Grok 4.3xAIArtificial Analysis τ²-bench | 97.7% | Not reported | Closed |
| 13 | DeepSeek V4 Pro 0813DeepSeekArtificial Analysis τ²-bench | 96.2% | Not reported | Open |
| 14 | Kimi K2.6Moonshot AIArtificial Analysis τ²-bench | 95.9% | Not reported | Open |
| 15 | Kimi K2.5 (Reasoning)Moonshot AIArtificial Analysis τ²-bench | 95.9% | Not reported | Closed |
| 16 | Kimi K2.5Moonshot AIArtificial Analysis τ²-bench | 95.9% | Not reported | Open |
| 17 | GLM-4.7Z.AIArtificial Analysis τ²-bench | 95.9% | Not reported | Open |
| 18 | Qwen 3.6 Max (preview)AlibabaArtificial Analysis τ²-bench | 95.9% | Not reported | Closed |
| 19 | Gemini 3.1 ProGoogleτ²-bench published setup | 95.6% | Not reported | Closed |
| 20 | Qwen3.5 397BAlibabaArtificial Analysis τ²-bench | 95.6% | Not reported | Open |
| 21 | Qwen3.5 397B (Reasoning)AlibabaArtificial Analysis τ²-bench | 95.6% | Not reported | Open |
| 22 | Gemini 3.5 FlashGoogleArtificial Analysis τ²-bench | 95.3% | Not reported | Closed |
| 23 | Qwen3.6-35B-A3BAlibabaArtificial Analysis τ²-bench | 95.3% | Not reported | Open |
| 24 | MiniMax M2.5MiniMaxArtificial Analysis τ²-bench | 95.3% | Not reported | Closed |
| 25 | MiMo-V2-ProXiaomiArtificial Analysis τ²-bench | 95% | Not reported | Closed |
| 26 | Qwen3.7 MaxAlibabaArtificial Analysis τ²-bench | 94.7% | Not reported | Closed |
| 27 | Claude Opus 4.8AnthropicArtificial Analysis τ²-bench | 94.4% | Not reported | Closed |
| 28 | Qwen3.6-27BAlibabaArtificial Analysis τ²-bench | 94.2% | Not reported | Open |
| 29 | MiMo-V2.5-ProXiaomiArtificial Analysis τ²-bench | 94.2% | Not reported | Closed |
| 30 | Mistral Medium 3.5 128BMistralArtificial Analysis τ²-bench | 94.2% | Not reported | Open |
| 31 | DeepSeek V4 Pro (High)DeepSeekArtificial Analysis τ²-bench | 94.2% | Not reported | Open |
| 32 | Qwen3.5-27BAlibabaArtificial Analysis τ²-bench | 93.9% | Not reported | Open |
| 33 | Qwen3.5-122B-A10BAlibabaArtificial Analysis τ²-bench | 93.6% | Not reported | Open |
| 34 | GPT-5.4 miniOpenAIτ²-bench Telecom | 93.4% | Not reported | Closed |
| 35 | Grok 4.1 Fast (Reasoning)xAIArtificial Analysis τ²-bench | 93.3% | Not reported | Closed |
| 36 | Qwen3.7 PlusAlibabaArtificial Analysis τ²-bench | 93% | Not reported | Closed |
| 37 | GPT-5.4 nanoOpenAIτ²-bench Telecom | 92.5% | Not reported | Closed |
| 38 | GPT-5.2-CodexOpenAIArtificial Analysis τ²-bench | 92.1% | Not reported | Closed |
| 39 | Claude Opus 4.6 (Adaptive)AnthropicArtificial Analysis τ²-bench | 92.1% | Not reported | Closed |
| 40 | Qwen3.5-4BAlibabaArtificial Analysis τ²-bench | 92.1% | Not reported | Open |
| 41 | Muse SparkMetaτ²-bench published setup | 91.5% | Not reported | Closed |
| 42 | MiMo-V2-OmniXiaomiArtificial Analysis τ²-bench | 91.2% | Not reported | Closed |
| 43 | Kimi K2.7 CodeMoonshot AIArtificial Analysis τ²-bench | 90.1% | Not reported | Open |
| 44 | Trinity-Large-ThinkingArcee AIArtificial Analysis τ²-bench | 90.1% | Not reported | Open |
| 45 | Trinity-Large-PreviewArcee AIArtificial Analysis τ²-bench | 90.1% | Not reported | Open |
| 46 | Claude Opus 4.5 ThinkingAnthropicArtificial Analysis τ²-bench | 89.5% | Not reported | Closed |
| 47 | Qwen3.5-35B-A3BAlibabaArtificial Analysis τ²-bench | 89.2% | Not reported | Open |
| 48 | MiniMax M3MiniMaxArtificial Analysis τ²-bench | 88.9% | Not reported | Open |
| 49 | Claude Opus 4.7 (Adaptive)AnthropicArtificial Analysis τ²-bench | 88.6% | Not reported | Closed |
| 50 | LFM2.5-8B-A1BLiquidAIτ²-bench Telecom | 88.1% | Not reported | Open |
| 51 | Step 3.5 FlashStepFunArtificial Analysis τ²-bench | 87.4% | Not reported | Open |
| 52 | Gemini 3 ProGoogleArtificial Analysis τ²-bench | 87.1% | Not reported | Closed |
| 53 | Spark-X2.6-4BSparkLLMτ²-bench published setup | 86.9% | Not reported | Open |
| 54 | GPT-5 (medium)OpenAIArtificial Analysis τ²-bench | 86.5% | Not reported | Closed |
| 55 | GPT-5.6 TerraOpenAIArtificial Analysis τ²-bench | 86.3% | Not reported | Closed |
| 56 | Claude Opus 4.5AnthropicArtificial Analysis τ²-bench | 86.3% | Not reported | Closed |
| 57 | Solar Pro 3UpstageArtificial Analysis τ²-bench | 86.3% | Not reported | Closed |
| 58 | GPT-5.3 CodexOpenAIArtificial Analysis τ²-bench | 86% | Not reported | Closed |
| 59 | Ling 2.6 FlashInclusionAIArtificial Analysis τ²-bench | 86% | Not reported | Open |
| 60 | GPT-5.3-Codex-SparkOpenAIArtificial Analysis τ²-bench | 86% | Not reported | Closed |
| 61 | GPT-5.6 SolOpenAIArtificial Analysis τ²-bench | 85.1% | Not reported | Closed |
| 62 | Command A+Cohereτ²-bench Telecom | 85% | Not reported | Open |
| 63 | Claude Opus 4.6AnthropicArtificial Analysis τ²-bench | 84.8% | Not reported | Closed |
| 64 | MiniMax M2.7MiniMaxArtificial Analysis τ²-bench | 84.8% | Not reported | Open |
| 65 | GPT-5.2OpenAIArtificial Analysis τ²-bench | 84.8% | Not reported | Closed |
| 66 | GPT-5 (high)OpenAIArtificial Analysis τ²-bench | 84.8% | Not reported | Closed |
| 67 | MiMo-V2-FlashXiaomiArtificial Analysis τ²-bench | 83.9% | Not reported | Open |
| 68 | Nemotron 3 UltraNVIDIAArtificial Analysis τ²-bench | 83.3% | Not reported | Open |
| 69 | GPT-5.1-CodexOpenAIArtificial Analysis τ²-bench | 83% | Not reported | Closed |
| 70 | GPT-5.1-Codex-MaxOpenAIArtificial Analysis τ²-bench | 83% | Not reported | Closed |
| 71 | GPT-5.1OpenAIArtificial Analysis τ²-bench | 81.9% | Not reported | Closed |
| 72 | o3OpenAIArtificial Analysis τ²-bench | 80.7% | Not reported | Closed |
| 73 | LLaDA2.2-flashInclusionAIτ²-bench published setup | 80.3% | Not reported | Open |
| 74 | Ternary Bonsai 2 27BPrism MLτ²-bench Telecom | 80.2% | Not reported | Open |
| 75 | Claude Sonnet 4.6AnthropicArtificial Analysis τ²-bench | 79.5% | Not reported | Closed |
| 76 | DeepSeek V3.2DeepSeekArtificial Analysis τ²-bench | 78.9% | Not reported | Open |
| 77 | Agents-A1-4BInternScienceτ²-bench published setup | 78.2% | Not reported | Open |
| 78 | GLM-4.6Z.AIArtificial Analysis τ²-bench | 76.9% | Not reported | Open |
| 79 | Grok Code Fast 1xAIArtificial Analysis τ²-bench | 75.7% | Not reported | Closed |
| 80 | Grok 4xAIArtificial Analysis τ²-bench | 74.9% | Not reported | Closed |
| 81 | Qwen3 MaxAlibabaArtificial Analysis τ²-bench | 74.3% | Not reported | Closed |
| 82 | K-ExaoneLG AI ResearchArtificial Analysis τ²-bench | 74.3% | Not reported | Closed |
| 83 | Claude Opus 4.7AnthropicArtificial Analysis τ²-bench | 74% | Not reported | Closed |
| 84 | Claude 4 Opus ThinkingAnthropicArtificial Analysis τ²-bench | 73.4% | Not reported | Closed |
| 85 | Claude 4.1 Opus ThinkingAnthropicArtificial Analysis τ²-bench | 71.4% | Not reported | Closed |
| 86 | Mercury 2InceptionArtificial Analysis τ²-bench | 70.8% | Not reported | Closed |
| 87 | GPT-5 miniOpenAIArtificial Analysis τ²-bench | 68.4% | Not reported | Closed |
| 88 | GPT-OSS 120BOpenAIArtificial Analysis τ²-bench | 65.8% | Not reported | Open |
| 89 | Grok 4 Fast (Reasoning)xAIArtificial Analysis τ²-bench | 65.8% | Not reported | Closed |
| 90 | Grok 4.1 FastxAIArtificial Analysis τ²-bench | 63.7% | Not reported | Closed |
| 91 | o1OpenAIArtificial Analysis τ²-bench | 62.6% | Not reported | Closed |
| 92 | Kimi K2Moonshot AIArtificial Analysis τ²-bench | 61.1% | Not reported | Closed |
| 93 | GPT-OSS 20BOpenAIArtificial Analysis τ²-bench | 60.2% | Not reported | Open |
| 94 | Gemma 4 31BGoogleArtificial Analysis τ²-bench | 59.9% | Not reported | Open |
| 95 | LLaDA2.2-miniInclusionAIτ²-bench published setup | 57.5% | Not reported | Open |
| 96 | o4-miniOpenAIArtificial Analysis τ²-bench | 55.6% | Not reported | Closed |
| 97 | Nemotron 3.5 Bolt E4B BF16NVIDIAτ²-bench published setup | 55.1% | Not reported | Open |
| 98 | Claude 3.7 Sonnet ThinkingAnthropicArtificial Analysis τ²-bench | 54.7% | Not reported | Closed |
| 99 | Gemini 2.5 ProGoogleArtificial Analysis τ²-bench | 54.1% | Not reported | Closed |
| 100 | GPT-4.1 miniOpenAIArtificial Analysis τ²-bench | 52.9% | Not reported | Closed |
| 101 | Nemotron 3.5 Bolt E4BNVIDIAτ²-bench published setup | 52.3% | Not reported | Open |
| 102 | Claude 4 SonnetAnthropicArtificial Analysis τ²-bench | 52.3% | Not reported | Closed |
| 103 | Claude 3.7 SonnetAnthropicArtificial Analysis τ²-bench | 50% | Not reported | Closed |
| 104 | GPT-4.1OpenAIArtificial Analysis τ²-bench | 47.1% | Not reported | Closed |
| 105 | DeepSeek V3 0324DeepSeekArtificial Analysis τ²-bench | 47.1% | Not reported | Open |
| 106 | Sarvam 105BSarvamArtificial Analysis τ²-bench | 46.8% | Not reported | Open |
| 107 | GLM-4.5-AirZ.AIArtificial Analysis τ²-bench | 46.5% | Not reported | Closed |
| 108 | Nemotron 3 Nano Omni 30B A3BNVIDIAτ²-bench Telecom | 45.3% | Not reported | Open |
| 109 | Gemma 4 26B A4BGoogleArtificial Analysis τ²-bench | 43.6% | Not reported | Open |
| 110 | Gemini 3 FlashGoogleArtificial Analysis τ²-bench | 43.3% | Not reported | Closed |
| 111 | Mistral Small 4MistralArtificial Analysis τ²-bench | 41.2% | Not reported | Open |
| 112 | DeepSeek V3.1 (Reasoning)DeepSeekArtificial Analysis τ²-bench | 37.4% | Not reported | Open |
| 113 | North Mini CodeCohereArtificial Analysis τ²-bench | 37.4% | Not reported | Open |
| 114 | DeepSeek-R1DeepSeekArtificial Analysis τ²-bench | 36.5% | Not reported | Open |
| 115 | GPT-5 nanoOpenAIArtificial Analysis τ²-bench | 36.5% | Not reported | Closed |
| 116 | Gemma 4 12BGoogleArtificial Analysis τ²-bench | 36.3% | Not reported | Open |
| 117 | DeepSeek V3.1DeepSeekArtificial Analysis τ²-bench | 34.8% | Not reported | Open |
| 118 | Sarvam 30BSarvamArtificial Analysis τ²-bench | 34.5% | Not reported | Open |
| 119 | MiniMax M1 80kMiniMaxArtificial Analysis τ²-bench | 34.2% | Not reported | Closed |
| 120 | Solar Pro 2UpstageArtificial Analysis τ²-bench | 31.9% | Not reported | Closed |
| 121 | Mistral Large 2MistralArtificial Analysis τ²-bench | 30.7% | Not reported | Closed |
| 122 | Gemini 2.0 FlashGoogleArtificial Analysis τ²-bench | 29.5% | Not reported | Closed |
| 123 | o3-miniOpenAIArtificial Analysis τ²-bench | 28.7% | Not reported | Closed |
| 124 | Devstral SmallMistralArtificial Analysis τ²-bench | 28.4% | Not reported | Open |
| 125 | Ministral 3 14B (Reasoning)MistralArtificial Analysis τ²-bench | 27.2% | Not reported | Open |
| 126 | Ministral 3 14BMistralArtificial Analysis τ²-bench | 27.2% | Not reported | Open |
| 127 | Llama 3.3 70BMetaArtificial Analysis τ²-bench | 26.6% | Not reported | Open |
| 128 | Magistral SmallMistralArtificial Analysis τ²-bench | 26.6% | Not reported | Open |
| 129 | Ultravox v0.6 Llama 3.3 70BFixie AIArtificial Analysis τ²-bench | 26.6% | Not reported | Open |
| 130 | Ministral 3 8B (Reasoning)MistralArtificial Analysis τ²-bench | 26.6% | Not reported | Open |
| 131 | Ministral 3 8BMistralArtificial Analysis τ²-bench | 26.6% | Not reported | Open |
| 132 | Nemotron 3 Nano 30BNVIDIAArtificial Analysis τ²-bench | 25.4% | Not reported | Open |
| 133 | GPT-4oOpenAIArtificial Analysis τ²-bench | 25.1% | Not reported | Closed |
| 134 | Mistral Small 3.1MistralArtificial Analysis τ²-bench | 25.1% | Not reported | Open |
| 135 | Ministral 3 3B (Reasoning)MistralArtificial Analysis τ²-bench | 24.9% | Not reported | Open |
| 136 | Ministral 3 3BMistralArtificial Analysis τ²-bench | 24.9% | Not reported | Open |
| 137 | Mistral Large 3MistralArtificial Analysis τ²-bench | 24.6% | Not reported | Closed |
| 138 | Claude 3.5 HaikuAnthropicArtificial Analysis τ²-bench | 24.6% | Not reported | Closed |
| 139 | Mistral Medium 3MistralArtificial Analysis τ²-bench | 24.3% | Not reported | Closed |
| 140 | Magistral MediumMistralArtificial Analysis τ²-bench | 23.1% | Not reported | Closed |
| 141 | DeepSeek V3DeepSeekArtificial Analysis τ²-bench | 22.8% | Not reported | Open |
| 142 | Llama 3.1 70BMetaArtificial Analysis τ²-bench | 21.6% | Not reported | Open |
| 143 | Qwen3-Omni-30B-A3B-ThinkingAlibabaArtificial Analysis τ²-bench | 21.3% | Not reported | Open |
| 144 | Claude 3 HaikuAnthropicArtificial Analysis τ²-bench | 21.1% | Not reported | Closed |
| 145 | Gemma 4 E4BGoogleArtificial Analysis τ²-bench | 20.8% | Not reported | Open |
| 146 | Gemma 4 E2BGoogleArtificial Analysis τ²-bench | 20.8% | Not reported | Open |
| 147 | Exaone 4.0 1.2BLG AI ResearchArtificial Analysis τ²-bench | 20.5% | Not reported | Open |
| 148 | Granite-4.0-H-1BIBMArtificial Analysis τ²-bench | 19.6% | Not reported | Open |
| 149 | LFM2.5-1.2B-ThinkingLiquidAIArtificial Analysis τ²-bench | 19.6% | Not reported | Closed |
| 150 | Gemini 2.5 Flash-LiteGoogleArtificial Analysis τ²-bench | 19% | Not reported | Closed |
| 151 | Llama 3.1 405BMetaArtificial Analysis τ²-bench | 19% | Not reported | Open |
| 152 | Mistral Small 4 (Reasoning)MistralArtificial Analysis τ²-bench | 18.4% | Not reported | Open |
| 153 | Llama 4 MaverickMetaArtificial Analysis τ²-bench | 17.8% | Not reported | Open |
| 154 | GPT-4.1 nanoOpenAIArtificial Analysis τ²-bench | 17.3% | Not reported | Closed |
| 155 | Qwen3-Omni-30B-A3B-InstructAlibabaArtificial Analysis τ²-bench | 16.4% | Not reported | Open |
| 156 | Llama 4 ScoutMetaArtificial Analysis τ²-bench | 15.5% | Not reported | Open |
| 157 | Gemini 2.5 FlashGoogleArtificial Analysis τ²-bench | 14.9% | Not reported | Closed |
| 158 | Granite-4.0-H-350MIBMArtificial Analysis τ²-bench | 14.6% | Not reported | Open |
| 159 | Nova ProAmazonArtificial Analysis τ²-bench | 14% | Not reported | Closed |
| 160 | Granite-4.0-350MIBMArtificial Analysis τ²-bench | 13.2% | Not reported | Open |
| 161 | Nemotron Ultra 253BNVIDIAArtificial Analysis τ²-bench | 11.4% | Not reported | Open |
| 162 | LFM2-24B-A2BLiquidAIArtificial Analysis τ²-bench | 11.1% | Not reported | Closed |
| 163 | LFM2.5-1.2B-InstructLiquidAIArtificial Analysis τ²-bench | 10.8% | Not reported | Closed |
| 164 | Gemma 3 27BGoogleArtificial Analysis τ²-bench | 10.5% | Not reported | Open |
| 165 | LFM2.5-VL-1.6B-ExtractLiquidAIArtificial Analysis τ²-bench | 8.5% | Not reported | Open |
| 166 | Exaone 4.0 32BLG AI ResearchArtificial Analysis τ²-bench | 4.1% | Not reported | Open |
| 167 | Phi-4MicrosoftArtificial Analysis τ²-bench | 0% | Not reported | Open |
| 168 | Phi-4-reasoningMicrosoftArtificial Analysis τ²-bench | 0% | Not reported | Open |
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.
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.
Freshness and provenance
Version
τ²-Bench 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
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
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.