Which AI model scores highest on τ²-bench?

GPT-5.1 by OpenAI holds the best τ²-bench (airline) result among tracked models, at 67% (released Nov 12 2025). Higher scores are better on this benchmark.

What are the top 4 models on τ²-bench?

1. GPT-5.1 (OpenAI) — 67%; 2. GPT-5 (OpenAI) — 62.6%; 3. GPT-5 mini (OpenAI) — 60%; 4. GPT-5 nano (OpenAI) — 41%.

How many models have a published τ²-bench score?

4 tracked models have a published τ²-bench (airline) score. Scores are collected from published lab reports and benchmark sources, with recorded sources credited alongside the results. We do not re-run the evaluations.

All benchmarks

Coding

ProgramBench SWE-Bench Pro SWE-Bench Verified SWE-Bench Multilingual SWE-Bench Multimodal CursorBench 4.0 DeepSWE 1.1 DeepSWE 1.0 SWE-sweep FrontierSWE V2 SWE-Marathon v1.1 FrontierCode v1.1 (Main)main split FrontierCode v1.1 (Extended)extended split FrontierCode 2.0 APEX-SWE MLE-Bench PaperBench Next.js Evals Supabase Evalswith skills Supabase Evalsno skills NL2Repo-Bench SWEAtlas CodeBase QnA QwenSWEBench QwenSWEBench V2 AA Coding Agent Index Database Migration Tasks (OpenAI Internal)CADGenBench BenchCAD BenchCADwithout tools BenchCADwith tools LiveCodeBench Pro LiveCodeBench Aider Polyglot (diff)HumanEval Expert-SWE (Internal)

Terminal & CLI

Frontier-Bench v0.1 Terminal-Bench 4.0 Terminal-Bench 3.0 Terminal-Bench 2.1 Terminal-Bench 2.0 Terminal-Bench-Science 0.1

Agentic & tool use

APEX-Agents MCP Atlas MCP Atlasoriginal harness JobBench CoWorkBench Toolathlon-Verified IFEval Advanced Cybersecurity Completion Rate (Internal)τ²-benchretail τ²-benchairline τ²-benchtelecom τ²-benchtelecom, adjusted prompt Toolathlon BU Bench BrowseComp DeepSearchQA CyberGym CVE-Bench CyScenarioBench CVE Rediscovery RoboDojo Sim WANDR (relaxed score)CathedralBench ExploitBench ExploitGym6-hour budget ExploitGym2-hour budget Agents’ Last Exam v1.1 CWE-bench v1 ExploitGym (causal pass rate)10-hour budget ExploitGym (causal pass rate)6-hour budget ExploitGym (remote target)ExploitGym Hardened RemoteExploitBench ExploitGym Honeypotwithout deployment prompt ExploitGym Honeypotwith deployment prompt SEC-bench Pro (JavaScript)OSWorld 2.0 OSWorld 2.1full set OSWorld 2.1offline OSWorld-Verified Agent's Last Exampass@1 Agent's Last Examscore SRE-Bench AutomationBench AutomationBench (Artificial Analysis)

Reasoning & science

De novo binder design (24 hr)RiemannBench Humanity's Last Examno tools Humanity's Last Examtext only, no tools Humanity's Last Examwith tools Humanity's Last Exam (Verified)ARC-AGI-3 ARC-AGI-2 FrontierMathTier 1–3 FrontierMathTier 4 FrontierMathTier 4 (v2)BioMysteryBenchhard BioMysteryBenchhuman solved LatchBio Capabilities v1.0 GeneBench-Pro MedChemBench (Internal)LAB-Bench 2 EEBench MATH MATH-500 DROP ARC-Challenge HellaSwag WinoGrande ARC-AGI-1 GeneBench LifeSciBench AIME 2024 AIME 2026 AIME 2025 FrontierMathTiers 1–3 (v2)GPQA Diamond IFBench Agentic IF Index (Internal)MMLU Okapi Multilingual MMLU MGSM GSM8K HLE-Diamond (no tools)

Knowledge work

SimpleQA Verified SimpleQA MMMLU MMLU-Pro AA Intelligence Index v5.0 Vals Index v2.2 Vending-Bench 2 (USD)AA Intelligence Index GDPval-AA GDPval-AA v2.1 GDPval-AA v2 Design Tasks (OpenAI Internal)Data Science Tasks (OpenAI Internal)AA-Briefcase v1.1 AA-Briefcase v1.1 (rubric)Remote Labor Index (RLI)AA-Briefcase GDPval (win/tie rate)OfficeQA Proextracted text

Finance

FinanceAgent v1.1 Finance Agent v2

Legal & tax

Harvey's Legal Agent Benchmark (Held-out)Harvey's Legal Agent Benchmark TaxEval v2

Healthcare

HealthBenchfull set HealthBenchfull set, length-adjusted HealthBenchhard HealthBenchhard, length-adjusted HealthBenchconsensus HealthBenchconsensus, length-adjusted HealthBench Professional HealthBench Professionalraw HealthBench ProfessionalOpenAI grading, raw HealthBench ProfessionalOpenAI grading, length-adjusted HealthBench Professionallength-adjusted MedScribe

Multimodal

CharXiv Reasoning CharXiv Reasoningwith tools Chartographywith tools Chartographyno tools OfficeQA Pro MVBench MMVU GDP.PDF GDP.PDF (Artificial Analysis)LVBench LVBenchagentic ZeroBench (tools, pass@5)SimpleVQA (with tools)BabyVision (with tools)BabyVision OpenScore String Quartets MathVista MMMU-Prostandard and vision mean MMMU-Prowith tools MMMU-Pro MMMU Blueprint-Bench 2

Long context

GraphWalks BFS (<128K)GraphWalks parents (<128K)OpenAI MRCR (2 needles)128K OpenAI MRCR (2 needles)256K OpenAI MRCR (2 needles)1M MRCR v2 (8-needle)256K–512K MRCR v2 (8-needle)512K–1M GraphWalks BFS F1 (up to 128K)GraphWalks BFS (256K–1M)MRCR v2 (8-needle)128k average MRCR v2 (8-needle)1M pointwise MRCR256k-512k MRCR512k-1M

Robustness

CheatBench BullshitBench v2 Gray Swan IPIk = 1 Gray Swan IPIk = 10 Gray Swan IPIk = 15 Auto-review circumvention (Internal)OpenAI Factuality (Internal)

Community preference

threejseval

Agentic tasks

τ²-benchairline

task success in the airline domain of the second-generation customer-service evaluation

retail airline telecom telecom, adjusted prompt

Rankings

Higher is better

τ²-bench — frequently asked questions

: task success in the airline domain of the second-generation customer-service evaluation
: GPT-5.1 by OpenAI holds the best τ²-bench (airline) result among tracked models, at 67% (released Nov 12 2025). Higher scores are better on this benchmark.
: 1. GPT-5.1 (OpenAI) — 67%; 2. GPT-5 (OpenAI) — 62.6%; 3. GPT-5 mini (OpenAI) — 60%; 4. GPT-5 nano (OpenAI) — 41%.
: 4 tracked models have a published τ²-bench (airline) score. Scores are collected from published lab reports and benchmark sources, with recorded sources credited alongside the results. We do not re-run the evaluations.