τ²-Bench Telecom
Customer-service agent benchmark in which both the agent and a simulated user act on a shared environment; this entry is the telecom domain.
Actively evaluated by Artificial Analysis on current frontier models; no saturation evidence recorded.
Performance Timeline
Longitudinal progression of model scores against human baselines.Performance & Historical Trajectory
Empirical score progression across model release dates and evaluation rounds.
| Model | Score | Date | Source Type | Provenance |
|---|---|---|---|---|
| JT-35B-Flash | 99.1% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 Omni Plus | 88.3% | 2026-11-15 | independent | Source ↗ |
| Nova 2.0 Omni | 80.4% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5-122B-A10B | 84.5% | 2026-11-15 | independent | Source ↗ |
| Gemma 4 12B | 31.9% | 2026-11-15 | independent | Source ↗ |
| Gemma 3 270M | 9.1% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 2B | 81.6% | 2026-11-15 | independent | Source ↗ |
| Qwen3 Omni 30B A3B Instruct | 16.4% | 2026-11-15 | independent | Source ↗ |
| Nova 2.0 Pro | 92.7% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 397B A17B | 95.6% | 2026-11-15 | independent | Source ↗ |
| Mistral Small 4 | 41.2% | 2026-11-15 | independent | Source ↗ |
| gpt-oss-20b | 50.3% | 2026-11-15 | independent | Source ↗ |
| o3 | 80.7% | 2026-11-15 | independent | Source ↗ |
| North Mini Code | 37.4% | 2026-11-15 | independent | Source ↗ |
| Granite 4.0 Micro | 12.6% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5-9B | 85.1% | 2026-11-15 | independent | Source ↗ |
| Hermes 4 - Llama-3.1 70B | 21.6% | 2026-11-15 | independent | Source ↗ |
| Nemotron 3 Nano 30B A3B | 25.4% | 2026-11-15 | independent | Source ↗ |
| Qwen3 Next 80B A3B Instruct | 21.6% | 2026-11-15 | independent | Source ↗ |
| MiniMax M3 | 88.9% | 2026-11-15 | independent | Source ↗ |
| NVIDIA Nemotron Nano 9B V2 | 23.4% | 2026-11-15 | independent | Source ↗ |
| Command A+ | 80.7% | 2026-11-15 | independent | Source ↗ |
| EXAONE 4.0 32B | 4.1% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 4B | 87.7% | 2026-11-15 | independent | Source ↗ |
| Olmo 3 7B Think | 0% | 2026-11-15 | independent | Source ↗ |
| HyperCLOVA X SEED Think (32B) | 87.4% | 2026-11-15 | independent | Source ↗ |
| HyperNova 60B 2605 | 63.2% | 2026-11-15 | independent | Source ↗ |
| Nova Micro | 14% | 2026-11-15 | independent | Source ↗ |
| INTELLECT-3 | 26.6% | 2026-11-15 | independent | Source ↗ |
| Olmo 3.1 32B Think | 0% | 2026-11-15 | independent | Source ↗ |
| Nex-N2-Pro | 81.6% | 2026-11-15 | independent | Source ↗ |
| LFM2.5-VL-1.6B | 8.5% | 2026-11-15 | independent | Source ↗ |
| JT-MINI | 93% | 2026-11-15 | independent | Source ↗ |
| Molmo2-8B | 0% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 0.8B | 65.2% | 2026-11-15 | independent | Source ↗ |
| MiniCPM-V 4.6 1.3B | 87.7% | 2026-11-15 | independent | Source ↗ |
| Sarvam 105B | 46.8% | 2026-11-15 | independent | Source ↗ |
| Llama 3.1 Instruct 405B | 19% | 2026-11-15 | independent | Source ↗ |
| Ring-flash-2.0 | 0% | 2026-11-15 | independent | Source ↗ |
| LFM2.5-1.2B-Instruct | 10.8% | 2026-11-15 | independent | Source ↗ |
| Apertus 70B Instruct | 12.9% | 2026-11-15 | independent | Source ↗ |
| Doubao Seed Code | 58.2% | 2026-11-15 | independent | Source ↗ |
| Gemma 4 E4B | 26% | 2026-11-15 | independent | Source ↗ |
| Granite 4.0 350M | 13.2% | 2026-11-15 | independent | Source ↗ |
| Kimi Linear 48B A3B Instruct | 0% | 2026-11-15 | independent | Source ↗ |
| Solar Open 100B | 48.2% | 2026-11-15 | independent | Source ↗ |
| GPT-5.3-Codex | 86% | 2026-11-15 | independent | Source ↗ |
| LFM2.5-1.2B-Thinking | 19.6% | 2026-11-15 | independent | Source ↗ |
| K2 Think V2 | 25.4% | 2026-11-15 | independent | Source ↗ |
| Nemotron 3 Ultra | 83.3% | 2026-11-15 | independent | Source ↗ |
| Tri-21B-Think | 81% | 2026-11-15 | independent | Source ↗ |
| Granite 4.0 H Small | 17.3% | 2026-11-15 | independent | Source ↗ |
| Llama 3.2 Instruct 11B (Vision) | 14.6% | 2026-11-15 | independent | Source ↗ |
| LongCat Flash Lite | 79.5% | 2026-11-15 | independent | Source ↗ |
| Nova 2.0 Lite | 72.8% | 2026-11-15 | independent | Source ↗ |
| Motif-2-12.7B-Reasoning | 46.5% | 2026-11-15 | independent | Source ↗ |
| Qwen3.6 35B A3B | 85.1% | 2026-11-15 | independent | Source ↗ |
| Ministral 3 3B | 24.9% | 2026-11-15 | independent | Source ↗ |
| Gemma 4 26B A4B | 43.6% | 2026-11-15 | independent | Source ↗ |
| NVIDIA Nemotron Nano 12B v2 VL | 21.3% | 2026-11-15 | independent | Source ↗ |
| Qwen3.5 Omni Flash | 84.5% | 2026-11-15 | independent | Source ↗ |
| Magistral Small 1.2 | 27.8% | 2026-11-15 | independent | Source ↗ |
| Ministral 3 14B | 27.2% | 2026-11-15 | independent | Source ↗ |
| Gemma 4 E2B | 22.2% | 2026-11-15 | independent | Source ↗ |
| Sarvam 30B | 34.5% | 2026-11-15 | independent | Source ↗ |
| MiniCPM5-1B | 81% | 2026-11-15 | independent | Source ↗ |
| Jamba 1.7 Large | 13.5% | 2026-11-15 | independent | Source ↗ |
| Nanbeige4.1-3B | 21.6% | 2026-11-15 | independent | Source ↗ |
| Tiny Aya Global | 0% | 2026-11-15 | independent | Source ↗ |
| Exaone 4.0 1.2B | 20.5% | 2026-11-15 | independent | Source ↗ |
| Hermes 4 - Llama-3.1 405B | 22.2% | 2026-11-15 | independent | Source ↗ |
| Magistral Medium 1.2 | 52% | 2026-11-15 | independent | Source ↗ |
| Apertus 8B Instruct | 11.4% | 2026-11-15 | independent | Source ↗ |
| Nemotron 3 Nano Omni 30B A3B Reasoning | 45.3% | 2026-11-15 | independent | Source ↗ |
| Kimi K2.7 Code | 90.1% | 2026-11-15 | independent | Source ↗ |
| Olmo 3 7B Instruct | 12.6% | 2026-11-15 | independent | Source ↗ |
| Llama 4 Maverick | 17.8% | 2026-11-15 | independent | Source ↗ |
| Olmo 3.1 32B Instruct | 21.3% | 2026-11-15 | independent | Source ↗ |
| Ministral 3 8B | 26.6% | 2026-11-15 | independent | Source ↗ |
| Llama 3.3 Instruct 70B | 26.6% | 2026-11-15 | independent | Source ↗ |
| Nemotron Cascade 2 30B A3B | 53.2% | 2026-11-15 | independent | Source ↗ |
| NVIDIA Nemotron 3 Nano 4B | 28.1% | 2026-11-15 | independent | Source ↗ |
| Phi-4-mini | 8.2% | 2026-11-15 | independent | Source ↗ |
| K2-V2 | 27.8% | 2026-11-15 | independent | Source ↗ |
| ERNIE 5.0 Thinking | 83.9% | 2026-11-15 | independent | Source ↗ |
| Granite 4.0 H 350M | 14.6% | 2026-11-15 | independent | Source ↗ |
| Apriel-v1.6-15B-Thinker | 69.3% | 2026-11-15 | independent | Source ↗ |
| Qwen3.7 Plus | 93% | 2026-11-15 | independent | Source ↗ |
| Step3 VL 10B | 16.1% | 2026-11-15 | independent | Source ↗ |
| Llama 4 Scout | 15.5% | 2026-11-15 | independent | Source ↗ |
| ERNIE 4.5 300B A47B | 0% | 2026-11-15 | independent | Source ↗ |
| Molmo 7B-D | 0% | 2026-11-15 | independent | Source ↗ |
| Jamba 1.7 Mini | 12.6% | 2026-11-15 | independent | Source ↗ |
| Trinity Large Thinking | 90.1% | 2026-11-15 | independent | Source ↗ |
| Qwen3 Omni 30B A3B (Reasoning) | 21.3% | 2026-11-15 | independent | Source ↗ |
| Gemini 3.1 Pro | 95.6% | 2026-11-15 | independent | Source ↗ |
| Falcon-H1R-7B | 27.8% | 2026-11-15 | independent | Source ↗ |
| Qwen3 Next 80B A3B (Reasoning) | 41.5% | 2026-11-15 | independent | Source ↗ |
| gpt-oss-120b | 65.8% | 2026-11-15 | independent | Source ↗ |
| Gemma 4 31B | 59.9% | 2026-11-15 | independent | Source ↗ |
| Llama 3.1 Nemotron Ultra 253B v1 | 11.4% | 2026-11-15 | independent | Source ↗ |
| Granite 4.0 H 1B | 19.6% | 2026-11-15 | independent | Source ↗ |
| Reka Flash 3 | 0% | 2026-11-15 | independent | Source ↗ |
| Mistral Medium 3.5 | 94.2% | 2026-11-15 | independent | Source ↗ |
| Mi:dm K 2.5 Pro | 86.5% | 2026-11-15 | independent | Source ↗ |
| LFM2.5-8B-A1B | 16.1% | 2026-11-15 | independent | Source ↗ |
| Nova Premier | 38.3% | 2026-11-15 | independent | Source ↗ |
| Ring-2.6-1T | 92.4% | 2026-11-15 | independent | Source ↗ |
| Phi-4 | 0% | 2026-11-15 | independent | Source ↗ |
| Jamba Reasoning 3B | 15.8% | 2026-11-15 | independent | Source ↗ |
| LFM2 24B A2B | 11.1% | 2026-11-15 | independent | Source ↗ |
| KAT Coder Pro V2 | 89.5% | 2026-11-15 | independent | Source ↗ |
| Qwen3 Coder Next | 79.5% | 2026-11-15 | independent | Source ↗ |
Human Baseline & Difficulty Horizon
Calibrated human reference points, specialist benchmarks, and ceiling thresholds.No peer-reviewed human baseline has been published for this benchmark.
Metric & Scoring Methodology
Verification protocols, aggregation formulas, and specialized metric variants.task success (%)Dataset & Compute Cost
Evaluation volume, public availability, API pricing, and local hardware requirements.$5 – $20 USD for full benchmark evaluation run on frontier APIs.
1x NVIDIA RTX 4090 (24GB) or A100 (40GB/80GB) via vLLM / SGLang
How to Run & Reproduce
Standardized evaluation protocols, CLI commands, and reproducible runner templates.lm_eval --model hf --model_args pretrained=<model_path> --tasks tau2-bench-telecom --batch_size autoopencompass --datasets tau2-bench-telecom --models <model_config># Standard API Evaluation Loop
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": prompt}],
temperature=0.0,
)When publishing results for τ²-Bench Telecom, always report the exact prompt template, few-shot exemplar ordering, sampling temperature (temperature=0), maximum reasoning budget tokens, and the precise timestamped model snapshot ID.
Contamination & Memorization Analysis
Audit of pretraining exposure risks, memorization vectors, and refresh policies.Static fixed snapshot
Public on web / HuggingFace