τ³-Bench Tool-Agent-User Evaluation (τ³-bench results)
τ³-bench is the current evolution of Sierra's tool-agent-user framework, adding corrected task releases and newer knowledge and voice evaluation modes alongside airline, retail, and telecom.
Top models on τ³-bench results — October 28, 2026
As of October 28, 2026, Mercury 2.5 leads the τ³-bench results leaderboard with 96.0% , followed by Mistral Medium 3.5 128B (91.4%) and MiMo-V2.5-Pro (72.9%).
Mercury 2.5
Inception
τ³-bench Telecom
Mistral Medium 3.5 128B
Mistral
τ³-bench Telecom
MiMo-V2.5-Pro
Xiaomi
τ³-bench published setup
32 modelsAgentic2% of Agentic reference weightCurrentUpdated October 28, 2026
Leaderboard (32 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Mercury 2.5Inceptionτ³-bench Telecom | 96.0% | Not reported | Closed |
| 2 | Mistral Medium 3.5 128BMistralτ³-bench Telecom | 91.4% | Not reported | Open |
| 3 | MiMo-V2.5-ProXiaomiτ³-bench published setup | 72.9% | Not reported | Closed |
| 4 | Nemotron 3 UltraNVIDIAτ³-bench published average | 70.9% | Not reported | Open |
| 5 | Qwen3.6 PlusAlibabaτ³-bench published setup | 70.7% | Not reported | Closed |
| 6 | GLM-5.1Z.AIτ³-bench published setup | 70.6% | Not reported | Open |
| 7 | Claude Opus 4.5Anthropicτ³-bench published setup | 70.2% | Not reported | Closed |
| 8 | Qwen3.5 397BAlibabaτ³-bench published setup | 68.4% | Not reported | Open |
| 9 | LFM3-1.7BLiquidAIτ³-bench published average | 68.0% | Not reported | Open |
| 10 | MiniMax M2.7MiniMaxτ³-bench published setup | 67.6% | Not reported | Open |
| 11 | Qwen3.6-35B-A3BAlibabaτ³-bench published setup | 67.2% | Not reported | Open |
| 12 | Pokee-Isaac 28BPokee AIτ³-bench Telecom | 66.2% | Not reported | Closed |
| 13 | Kimi K2.5Moonshot AIτ³-bench published setup | 65.7% | Not reported | Open |
| 14 | GLM-5Z.AIτ³-bench published setup | 65.6% | Not reported | Open |
| 15 | Granite 4.2 30BIBMτ³-bench published setup | 62.0% | Not reported | Open |
| 16 | Granite 4.2 8BIBMτ³-bench published average | 58.1% | Not reported | Open |
| 17 | Ling 3.1 FlashInclusionAIτ³-bench published average | 47.8% | Not reported | Open |
| 18 | Granite 4.2 3BIBMτ³-bench published setup | 45.8% | Not reported | Open |
| 19 | Atria Dawn PreviewShanghai Artificial Intelligence Laboratoryτ³-bench published setup | 41.2% | Not reported | Open |
| 20 | Beam 1.0 501B-A23B FP8Reflection AIτ³-bench published setup | 36.8% | Not reported | Open |
| 21 | Beam 1.0 501B-A23BReflection AIτ³-bench published setup | 36.1% | Not reported | Open |
| 22 | Beam 1.0 501B-A23B NVFP4Reflection AIτ³-bench published setup | 35.4% | Not reported | Open |
| 23 | Spark-X2.6-4BSparkLLMτ³-bench published average | 34.7% | Not reported | Open |
| 24 | Mercury 3Inceptionτ³-bench Telecom | 30.2% | Not reported | Closed |
| 25 | Nemotron 3.5 Super VL 120B A12BNVIDIAτ³-bench published average | 18.9% | Not reported | Open |
| 26 | Nemotron 3.5 Super VL 120B A12B NVFP4NVIDIAτ³-bench published average | 17.8% | Not reported | Open |
| 27 | Nemotron 3.5 Bolt 74B-A9B BF16NVIDIAτ³-bench published setup | 17.0% | Not reported | Open |
| 28 | Nemotron 3.5 Bolt 74B-A9BNVIDIAτ³-bench published setup | 17.0% | Not reported | Open |
| 29 | Nemotron 3.5 Lightning 30B A3B NVFP4NVIDIAτ³-bench published setup | 9.5% | Not reported | Open |
| 30 | Mercury 2Inceptionτ³-bench published setup | 9.0% | Not reported | Closed |
| 31 | LFM2.5-24B-A2BLiquidAIτ³-bench published setup | 8.0% | Not reported | Open |
| 32 | LFM2.5-2.6BLiquidAIτ³-bench published setup | 5.7% | Not reported | Open |
How to read this leaderboard
Editorial review by Glevd · 2026-07-15
Use each result with its attached source and setup label. Match domain or published average, task release, modality, agent and user models, scaffold, prompts, trial count, and pass^k policy before comparison. The sorted table does not imply one controlled benchmark run.
Operator receipt: 32 sourced rows are currently displayable on this page; the highest published score among these 32 models is Mercury 2.5 at 96.0%, which does not establish a market leader.
Honest limit: Ten sourced rows come from several provider reports, and their labels include telecom and broader published averages. BenchLM did not rerun them. The framework's active task fixes and expanding modalities also mean an older result may not describe the current release.
According to BenchLM.ai, Mercury 2.5 leads the τ³-bench results benchmark with a score of 96.0%, followed by Mistral Medium 3.5 128B (91.4%) and MiMo-V2.5-Pro (72.9%). There is significant spread across the leaderboard, making this benchmark effective at differentiating model capabilities.
32 models have been evaluated on τ³-bench results. The benchmark falls in the Agentic category. BenchAlign v5.8 gives τ³-bench results 2% of the Agentic reference weight, so it moves the Agentic leaderboard and the overall ranking. Reference weights are relative weights in the calibrated model, not fixed shares of a score.
About τ³-bench results
Year
2026
Tasks
Corrected customer-service tasks plus knowledge and voice evaluation modes
Format
Published domain or average success results
Difficulty
Long-horizon, multimodal, and knowledge-aware tool use
The maintained repository now identifies the framework as τ³-bench and documents text, voice, telecom, airline, retail, and knowledge-aware banking evaluation. BenchLM keeps provider-published τ³ rows separate from original TAU-bench and τ²-bench rows.
Freshness and provenance
Version
τ³-bench results 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
Is τ³-bench the same as original TAU-bench?
No. It is the maintained successor framework with corrected tasks and newer domains and modalities. BenchLM keeps original TAU-bench, τ²-bench, τ² Airline, and τ³-bench on separate routes.
Can every τ³-bench result be ranked together?
No. Match the domain or average, release, modality, agent and user models, scaffold, prompts, trials, and pass^k policy. Provider tables without those same controls are useful source receipts, not one apples-to-apples leaderboard.
Compare top models on τ³-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.