{
 "body": "\n## What it measures\n\n\u03c4-bench measures whether a language agent can act as a customer-service representative: chat with a\nuser that is itself simulated by an LLM, use domain-specific API tools to look up and modify records,\nand follow a written policy document, inside a retail or an airline domain. Unlike a single-turn\ntool-call benchmark, the conversation is dynamic. The simulated user pursues its own goal and only\nreveals what it would naturally reveal, so the agent must ask questions rather than execute a scripted\ncall sequence. The modality is text, and interaction runs as multi-turn dialogue with structured tool\ncalls interleaved.\n\n## How it is scored\n\nA task succeeds if the database state at the end of the conversation matches an annotated goal state,\nnot if the transcript merely sounds right. The paper's pass^k metric asks, across n independent trials\nof the same task, what fraction of all size-k subsets of those trials are all-successes, which measures\nwhether an agent solves a task reliably rather than just once. pass^1 is the ordinary single-trial\nsuccess rate. Running at least 3 trials per task, GPT-4o with function calling succeeded on under 50% of\nretail tasks by pass^1 and under 25% by pass^8, showing a large drop in reliability as more repeated\ntrials are demanded.\n\n## Dataset and licence\n\nThe original release has 165 tasks: 115 in the retail domain and 50 in the airline domain, each with a\nnatural-language user goal and an annotated ground-truth database end-state, plus a shared policy\ndocument and tool set per domain. The github.com/sierra-research/tau-bench repository is MIT licensed.\nThere is no held-out answer key beyond the goal states shipped in the repository; everything needed to\ngrade a run is public.\n\n## Who publishes it\n\n\u03c4-bench comes from Sierra, an AI agent company, with the paper \"\u03c4-bench: A Benchmark for\nTool-Agent-User Interaction in Real-World Domains\" (June 2024) authored by Shunyu Yao, Noah Shinn, Pedram\nRazavi and Karthik Narasimhan. Sierra continues to lead its successors, and a public leaderboard for the\ncurrent generation is hosted at taubench.com per the tau2-bench repository's own documentation.\n\n## Lineage\n\nThe original \u03c4-bench repository now carries its own notice that its retail and airline tasks are not\nupdated, directing users to newer generations instead. \u03c4\u00b2-bench followed, adding a dual-control\nenvironment where the simulated user, not just the agent, can also take actions, plus telecom and a\nknowledge-retrieval-focused banking domain. \u03c4\u00b3-bench, announced by Sierra in March 2026, layers on a\nconfigurable retrieval-augmented knowledge domain (\u03c4-Knowledge, including a banking test with 698\ndocuments) and full-duplex voice evaluation (\u03c4-Voice), plus community-contributed fixes to the existing\ndomains. Neither \u03c4\u00b2-bench nor \u03c4\u00b3-bench has its own id or page in this repository yet; scores recorded\nunder this id should be assumed to be the original retail/airline task set unless a model's own\ndocumentation says otherwise.\n\n## Saturation and contamination\n\nAt publication, a strong function-calling agent (GPT-4o) cleared under half of retail tasks on a single\ntry and under a quarter when asked to repeat the same task 8 times successfully in a row, so the\noriginal paper treats reliability, not raw pass@1, as the open problem: open, as of June 2024. A current\ndated leaderboard reading for this specific original task set was not established, since attention\nappears to have shifted to \u03c4\u00b2/\u03c4\u00b3-bench. On contamination: task policies, tools and goal states have\nbeen public on GitHub since June 2024, so a model could plausibly have seen them in pretraining; the\nmulti-turn, simulated-user design limits how directly useful verbatim memorization is, but the authors\ndo not describe a specific mitigation.\n\n## How to run it\n\nThe reference harness is github.com/sierra-research/tau-bench for the original retail/airline tasks, or\ngithub.com/sierra-research/tau2-bench for the current generation. A trial requires both an agent model\nand a separate LLM to simulate the user, so a reported score depends on which model plays the user as\nwell as which plays the agent, a detail that is easy to omit and that changes results. None of the exact\ntask names were confirmed in lm-evaluation-harness, inspect_evals, HELM, OpenCompass or BIG-bench's\npublished task lists at the time of this research.\n\n## Reading the numbers\n\nA high \u03c4-bench score means an agent can hold a realistic multi-turn support conversation, extract the\nright information from a simulated customer, call the correct tools, and leave the backend database in\nthe state a human agent would have left it in, all while following written policy. Because of pass^k,\nlook at more than one trial count: a model with a good pass^1 but a weak pass^8 is inconsistent, not just\nimperfect, which matters more for production deployment than a single success number suggests. Check\nwhich task generation (original \u03c4-bench versus \u03c4\u00b2/\u03c4\u00b3-bench) and which user-simulator model produced a\ngiven score before comparing it to another report.\n",
 "build": {
  "built_at": "2026-09-09T16:56:50+00:00",
  "commit": "0a599558854c0e238c03a0f0d725239cb28f9d11",
  "eligibility_as_of": "2026-09-09"
 },
 "disposition": {
  "canonical_id": "tau_bench",
  "reasons": [],
  "status": "unassessed",
  "verified_results": []
 },
 "models_covered": [
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GLM 5.1",
   "model_id": "zhipu/glm-5-1",
   "provider": "zhipu",
   "provider_display": "Z.ai (Zhipu AI)",
   "score": 70.6,
   "source": "zai-org-model-card"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "Claude Opus 4",
   "model_id": "anthropic/claude-opus-4-20250514",
   "provider": "anthropic",
   "provider_display": "Anthropic",
   "score": 68.2,
   "source": "lmarena.ai, provider-reports, multimodal-evals, safety-evals, preference-evals, domain-evals"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "Claude Opus 4.6",
   "model_id": "anthropic/claude-opus-4-6",
   "provider": "anthropic",
   "provider_display": "Anthropic",
   "score": 68.2,
   "source": "lmarena.ai, provider-reports, multimodal-evals, safety-evals, preference-evals, domain-evals, anthropic-system-card-mythos"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "o3",
   "model_id": "openai/o3",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 65.5,
   "source": "lmarena.ai, provider-reports, domain-evals"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "o3-deep-research",
   "model_id": "openai/o3-deep-research",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 65.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "o3-mini",
   "model_id": "openai/o3-mini",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 65.5,
   "source": "lmarena.ai, provider-reports, llm-stats, domain-evals"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "o3-pro",
   "model_id": "openai/o3-pro",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 65.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "Claude Opus 4.1",
   "model_id": "anthropic/claude-opus-4-1-20250805",
   "provider": "anthropic",
   "provider_display": "Anthropic",
   "score": 64.8,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "Claude Opus 4.1 (latest)",
   "model_id": "anthropic/claude-opus-4-1",
   "provider": "anthropic",
   "provider_display": "Anthropic",
   "score": 64.8,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "Claude Sonnet 4",
   "model_id": "anthropic/claude-sonnet-4-20250514",
   "provider": "anthropic",
   "provider_display": "Anthropic",
   "score": 61.5,
   "source": "lmarena.ai, provider-reports, multimodal-evals, safety-evals, preference-evals"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "Claude Sonnet 4.5",
   "model_id": "anthropic/claude-sonnet-4-5-20250929",
   "provider": "anthropic",
   "provider_display": "Anthropic",
   "score": 61.5,
   "source": "lmarena.ai, provider-reports, multimodal-evals, safety-evals, preference-evals"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "Claude Sonnet 4.5 (latest)",
   "model_id": "anthropic/claude-sonnet-4-5",
   "provider": "anthropic",
   "provider_display": "Anthropic",
   "score": 61.5,
   "source": "lmarena.ai, provider-reports, multimodal-evals, safety-evals, preference-evals"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5",
   "model_id": "openai/gpt-5",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports, domain-evals"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5 Chat (latest)",
   "model_id": "openai/gpt-5-chat-latest",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5 Mini",
   "model_id": "openai/gpt-5-mini",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5 Nano",
   "model_id": "openai/gpt-5-nano",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5 Pro",
   "model_id": "openai/gpt-5-pro",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5-Codex",
   "model_id": "openai/gpt-5-codex",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.1",
   "model_id": "openai/gpt-5-1",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports, domain-evals"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.1 Chat",
   "model_id": "openai/gpt-5-1-chat-latest",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.1 Codex",
   "model_id": "openai/gpt-5-1-codex",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.1 Codex Max",
   "model_id": "openai/gpt-5-1-codex-max",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.1 Codex mini",
   "model_id": "openai/gpt-5-1-codex-mini",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.2",
   "model_id": "openai/gpt-5-2",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports, domain-evals"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.2 Chat",
   "model_id": "openai/gpt-5-2-chat-latest",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.2 Codex",
   "model_id": "openai/gpt-5-2-codex",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.2 Pro",
   "model_id": "openai/gpt-5-2-pro",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.3 Chat (latest)",
   "model_id": "openai/gpt-5-3-chat-latest",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.3 Codex",
   "model_id": "openai/gpt-5-3-codex",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.3 Codex Spark",
   "model_id": "openai/gpt-5-3-codex-spark",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.4",
   "model_id": "openai/gpt-5-4",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports, anthropic-system-card-mythos, domain-evals"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.4 mini",
   "model_id": "openai/gpt-5-4-mini",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.4 nano",
   "model_id": "openai/gpt-5-4-nano",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-5.4 Pro",
   "model_id": "openai/gpt-5-4-pro",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 58.5,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "Gemini 2.5 Pro",
   "model_id": "google/gemini-2-5-pro",
   "provider": "google",
   "provider_display": "Google DeepMind",
   "score": 58.2,
   "source": "lmarena.ai, provider-reports, multimodal-evals, safety-evals, preference-evals, domain-evals, llm-stats, intlpull"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "Qwen3-Coder 480B-A35B Instruct",
   "model_id": "qwen/qwen3-coder-480b-a35b-instruct",
   "provider": "qwen",
   "provider_display": "Alibaba / Qwen Team",
   "score": 55.8,
   "source": "lmarena.ai, provider-reports"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-4.1",
   "model_id": "openai/gpt-4-1",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 52.8,
   "source": "lmarena.ai, provider-reports, multimodal-evals, safety-evals, domain-evals preference-evals, llm-stats, intlpull"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "Qwen 3 235B Instruct",
   "model_id": "cerebras/qwen-3-235b-a22b-instruct-2507",
   "provider": "cerebras",
   "provider_display": "Cerebras",
   "score": 50.2,
   "source": "bigcode-leaderboard, provider-reports, preference-evals"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "Qwen3 235B-A22B",
   "model_id": "qwen/qwen3-235b-a22b",
   "provider": "qwen",
   "provider_display": "Alibaba / Qwen Team",
   "score": 50.2,
   "source": "lmarena.ai, provider-reports, preference-evals, llm-stats, domain-evals"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "DeepSeek R1",
   "model_id": "deepseek/deepseek-r1",
   "provider": "deepseek",
   "provider_display": "DeepSeek",
   "score": 48.8,
   "source": "lmarena.ai, provider-reports, preference-evals, open-llm-leaderboard-v2, domain-evals"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "DeepSeek R1 0528",
   "model_id": "deepseek/deepseek-r1-0528",
   "provider": "deepseek",
   "provider_display": "DeepSeek",
   "score": 48.8,
   "source": "lmarena.ai, provider-reports, preference-evals, open-llm-leaderboard-v2, domain-evals"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "DeepSeek R1 0528 NVFP4 v2",
   "model_id": "nvidia/deepseek-r1-0528-nvfp4-v2",
   "provider": "nvidia",
   "provider_display": "NVIDIA",
   "score": 48.8,
   "source": "lmarena.ai, provider-reports, preference-evals, open-llm-leaderboard-v2"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "DeepSeek Reasoner",
   "model_id": "deepseek/deepseek-reasoner",
   "provider": "deepseek",
   "provider_display": "DeepSeek",
   "score": 48.8,
   "source": "lmarena.ai, provider-reports, preference-evals, open-llm-leaderboard-v2, llm-stats"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-4o",
   "model_id": "openai/gpt-4o",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 42.5,
   "source": "lmarena.ai, provider-reports, multimodal-evals, safety-evals, preference-evals, domain-evals, llm-stats, intlpull"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-4o (2024-05-13)",
   "model_id": "openai/gpt-4o-2024-05-13",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 42.5,
   "source": "lmarena.ai, provider-reports, multimodal-evals, safety-evals, preference-evals, domain-evals, llm-stats, intlpull"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-4o (2024-08-06)",
   "model_id": "openai/gpt-4o-2024-08-06",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 42.5,
   "source": "lmarena.ai, provider-reports, multimodal-evals, safety-evals, preference-evals, domain-evals, llm-stats, intlpull"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-4o (2024-11-20)",
   "model_id": "openai/gpt-4o-2024-11-20",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 42.5,
   "source": "lmarena.ai, provider-reports, multimodal-evals, safety-evals, preference-evals, domain-evals, llm-stats, intlpull"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "GPT-4o mini",
   "model_id": "openai/gpt-4o-mini",
   "provider": "openai",
   "provider_display": "OpenAI",
   "score": 42.5,
   "source": "lmarena.ai, provider-reports, multimodal-evals, safety-evals, preference-evals, domain-evals, llm-stats, intlpull"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "DeepSeek Chat",
   "model_id": "deepseek/deepseek-chat",
   "provider": "deepseek",
   "provider_display": "DeepSeek",
   "score": 41.2,
   "source": "lmarena.ai, provider-reports, safety-evals, preference-evals, open-llm-leaderboard-v2, llm-stats"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "DeepSeek V3",
   "model_id": "deepseek/deepseek-v3",
   "provider": "deepseek",
   "provider_display": "DeepSeek",
   "score": 41.2,
   "source": "lmarena.ai, provider-reports, safety-evals, preference-evals, domain-evals open-llm-leaderboard-v2, llm-stats"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "DeepSeek V3 0324",
   "model_id": "deepseek/deepseek-v3-0324",
   "provider": "deepseek",
   "provider_display": "DeepSeek",
   "score": 41.2,
   "source": "lmarena.ai, provider-reports, safety-evals, preference-evals, open-llm-leaderboard-v2, llm-stats"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "DeepSeek V3.1",
   "model_id": "deepseek/deepseek-v3-1",
   "provider": "deepseek",
   "provider_display": "DeepSeek",
   "score": 41.2,
   "source": "lmarena.ai, provider-reports, safety-evals, preference-evals, open-llm-leaderboard-v2, llm-stats"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "DeepSeek V3.2",
   "model_id": "deepseek/deepseek-v3-2",
   "provider": "deepseek",
   "provider_display": "DeepSeek",
   "score": 41.2,
   "source": "lmarena.ai, provider-reports, safety-evals, preference-evals, open-llm-leaderboard-v2, llm-stats"
  },
  {
   "as_of": "2026-04",
   "attribution": "unverified-legacy",
   "display_name": "DeepSeek V3.2 Exp",
   "model_id": "deepseek/deepseek-v3-2-exp",
   "provider": "deepseek",
   "provider_display": "DeepSeek",
   "score": 41.2,
   "source": "lmarena.ai, provider-reports, safety-evals, preference-evals, open-llm-leaderboard-v2, llm-stats"
  }
 ],
 "page": {
  "aliases": [
   "tau-bench",
   "tau bench"
  ],
  "category": "agentic",
  "contamination": {
   "note": "Task policies, tool schemas and goal states have been public on GitHub since June 2024, so a model could have seen them in pretraining. The multi-turn, simulated-user design makes verbatim memorization less directly useful than for a static QA set, but the authors do not describe a specific mitigation.",
   "risk": "medium"
  },
  "dataset": {
   "languages": [
    "en"
   ],
   "license": "MIT",
   "modalities": [
    "text"
   ],
   "public_test_set": true,
   "size": 165,
   "size_note": "115 retail-domain tasks plus 50 airline-domain tasks in the original release, each with a natural-language user goal and an annotated database end-state.",
   "splits": "single evaluation set per domain; no train/test split",
   "url": "https://github.com/sierra-research/tau-bench"
  },
  "freshness": {
   "researched": "2026-09-07",
   "researched_by": "sonnet-5 agent, batch 1, slice H",
   "reviewed": "",
   "reviewed_by": ""
  },
  "harness": {
   "bigbench": "",
   "helm": "",
   "inspect_evals": "",
   "lm_eval": "",
   "opencompass": "",
   "other": ""
  },
  "id": "tau_bench",
  "last_updated": "",
  "leaderboard_url": "https://www.taubench.com",
  "lineage": {
   "family": "",
   "predecessor": "",
   "successors": [],
   "variants": []
  },
  "measures": "\u03c4-bench asks a language agent to act as a customer-service representative in a retail or an airline domain: it chats with a user that is itself simulated by an LLM, calls domain-specific API tools to look up and change records, and must follow a written policy document. The user pursues its own goal and only reveals what it would naturally reveal, so the agent has to ask questions rather than execute a fixed call sequence.",
  "metric": {
   "baseline_note": "pass^k is the fraction of all size-k subsets of a task's independent trials in which every trial succeeds, averaged across tasks; it measures reliability rather than a single success. pass^1 is the plain single-trial success rate. In the original paper (\u22653 trials per task), GPT-4o with function calling succeeded on under 50% of retail tasks by pass^1 and under 25% by pass^8.",
   "direction": "higher_is_better",
   "human_baseline": null,
   "max_score": 100.0,
   "name": "pass^k",
   "random_baseline": null,
   "unit": "%"
  },
  "name": "\u03c4-bench",
  "page_kind": "benchmark",
  "paper": {
   "arxiv": "2406.12045",
   "title": "\u03c4-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains",
   "url": "https://arxiv.org/abs/2406.12045",
   "year": 2024
  },
  "publisher": {
   "authors": [
    "Shunyu Yao",
    "Noah Shinn",
    "Pedram Razavi",
    "Karthik Narasimhan"
   ],
   "org": "Sierra",
   "url": "https://sierra.ai/"
  },
  "released": "2024-06",
  "repo_url": "https://github.com/sierra-research/tau-bench",
  "saturation": {
   "as_of": "2024-06",
   "note": "The original paper reports GPT-4o succeeding on under 50% of retail tasks by pass^1 and under 25% by pass^8, an inequality rather than a single exact top score, so no top_score is recorded here. A current, dated leaderboard reading for this original retail/airline task set specifically was not established; field attention appears to have moved to \u03c4\u00b2-bench and \u03c4\u00b3-bench.",
   "status": "open",
   "top_score": null
  },
  "sources": [
   {
    "accessed": "2026-09-07",
    "title": "\u03c4-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains (arXiv abstract)",
    "url": "https://arxiv.org/abs/2406.12045"
   },
   {
    "accessed": "2026-09-07",
    "title": "\u03c4-bench paper, HTML rendering (task counts, pass^k definition)",
    "url": "https://arxiv.org/html/2406.12045"
   },
   {
    "accessed": "2026-09-07",
    "title": "sierra-research/tau-bench GitHub repository",
    "url": "https://github.com/sierra-research/tau-bench"
   },
   {
    "accessed": "2026-09-07",
    "title": "sierra-research/tau2-bench GitHub repository",
    "url": "https://github.com/sierra-research/tau2-bench"
   },
   {
    "accessed": "2026-09-07",
    "title": "Sierra blog: \u03c4-bench, shaping the development and evaluation of agents",
    "url": "https://sierra.ai/blog/tau-bench-shaping-development-evaluation-agents"
   },
   {
    "accessed": "2026-09-07",
    "title": "Sierra blog: \u03c4\u00b3-bench, advancing agent benchmarking to knowledge and voice",
    "url": "https://sierra.ai/blog/bench-advancing-agent-benchmarking-to-knowledge-and-voice"
   }
  ],
  "status": "superseded",
  "subcategory": "tool-use customer-service agents",
  "summary": "Scores whether a tool-using agent can hold a policy-following conversation with a simulated customer and leave the backend in the right state.",
  "tags": [
   "tool-use",
   "multi-turn",
   "customer-service",
   "agentic",
   "simulated-user"
  ],
  "task_format": "multi-turn text dialogue interleaved with structured tool calls, against a simulated user model"
 }
}