Benchmark family

τ-bench

A family groups the versions of one benchmark. It is not evaluated and has no applicable models of its own; each of its 7 definitions is resolved separately.

Customer Service AgentKnowledge Grounded AgentMulti Participant AgentSoftware EngineeringTool UseVoice AgentAgent ConstructionTool Use

Definitions

Compare the newest 4 →

Newest first, following the recorded successor chain. Each definition is resolved against models separately.

Benchmarkτ-Knowledge / τ-Banking
ActiveNewest in family
Derived from τ²-bench
Related to τ³-bench v1.0.1
Scale
51 tools
Primary metric
Pass K Reliability
Applicable models
51
Benchmarkτ-Voice
Active
Derived from τ²-bench
Related to τ³-bench v1.0.1
Scale
50 counts
Primary metric
Pass K Reliability
Applicable models
9
Benchmarkτ^τ-bench / Hyper-τ
Active
Derived from τ³-bench v1.0.1
Scale
6 counts
Primary metric
Budget Adjusted Reward
Applicable models
6
Benchmarkτ³-bench v1.0.11.0.1
Active
Successor of τ³-bench v1.0.0
Scale
50 counts
Primary metric
Pass K Reliability
Applicable models
51
Benchmarkτ³-bench v1.0.01.0.0
Superseded
Successor of τ²-bench
Scale
50 counts
Primary metric
Pass K Reliability
Applicable models
51
Benchmarkτ²-bench
Superseded
Successor of τ-bench
Scale
50 counts
Primary metric
Pass K Reliability
Applicable models
51
Benchmarkτ-bench
Superseded
Scale
50 counts
Primary metric
Pass K Reliability
Applicable models
51

Family methodology

No family-level methodology recorded.

Across definitions

Tasks: End To End Agent Construction, Requirement Clarification, Dual Control Shared State Coordination, Policy Constrained Conversational Tool Use, Knowledge Grounded Conversational Tool Use, Full Duplex Voice Task Execution
Sources: None recorded at family level
Last reviewed: 2026-09-23

Applicability does not mean evaluated; benchmark result/model score data is outside this UI contract.