Benchmark leaderboard - 25 entries on 2 boards
tau-bench and tau2-bench leaderboard
tau-bench measures how well AI agents converse with users, call tools, retrieve knowledge, and follow policy across enterprise customer-service domains, in text and voice. As of October 17, 2026, Qwen3.5-397B-A17B (Alibaba Cloud) ranks first on the tau2-bench (Overall) board with 87.9% and Qwen 3.8 Max (Qwen) ranks first on the tau3-Banking (tau-knowledge) board with 55.2%.
Last verified from taubench.com. Read the caveats
This page covers 2 separate boards: tau2-bench (Overall) and tau3-Banking (tau-knowledge). Each is scored on its own terms, so compare rows only within the same board.
tau2-bench (Overall)
8 entries, in the board's order
| 1 | Qwen3.5-397B-A17B (Alibaba Cloud) tau2-bench leaderboard, Overall pass^1 = mean of retail 84.43 / airline 81.5 / telecom 97.81; standard submission by Sierra; reasoning effort: enabled; user simulator gpt-5.2; verified; rank #1 of 8 Official leaderboard | 87.9% |
| 2 | Claude Opus 4.5 (Anthropic) tau2-bench leaderboard, Overall pass^1 = mean of retail 79.61 / airline 84 / telecom 92.32; standard submission by Sierra; reasoning effort: high; user simulator gpt-5.2; verified; rank #2 of 8 Official leaderboard | 85.3% |
| 3 | GPT-5.2 (OpenAI) tau2-bench leaderboard, Overall pass^1 = mean of retail 81.58 / airline 83 / telecom 89.69; standard submission by Sierra; reasoning effort: high; user simulator gpt-5.2; verified; rank #3 of 8 Official leaderboard | 84.8% |
| 4 | Gemini 3 Flash (Google) tau2-bench leaderboard, Overall pass^1 = mean of retail 76.75 / airline 82.5 / telecom 91.23; standard submission by Sierra; reasoning effort: high; user simulator gpt-5.2; verified; rank #4 of 8 Official leaderboard | 83.5% |
| 5 | Gemini 3 Pro (Google) tau2-bench leaderboard, Overall pass^1 = mean of retail 75.88 / airline 80.5 / telecom 91.01; standard submission by Sierra; reasoning effort: high; user simulator gpt-5.2; verified; rank #5 of 8 Official leaderboard | 82.5% |
| 6 | GLM-5 (Zhipu AI) tau2-bench leaderboard, Overall pass^1 = mean of retail 73.68 / airline 82.5 / telecom 86.84; standard submission by Sierra; reasoning effort: enabled; user simulator gpt-5.2; verified; rank #6 of 8 Official leaderboard | 81.0% |
| 7 | Claude Sonnet 4.5 (Anthropic) tau2-bench leaderboard, Overall pass^1 = mean of retail 72.37 / airline 72 / telecom 84.87; standard submission by Sierra; reasoning effort: enabled; user simulator gpt-5.2; verified; rank #7 of 8 Official leaderboard | 76.4% |
| 8 | GPT-5.2 (OpenAI) tau2-bench leaderboard, Overall pass^1 = mean of retail 75 / airline 52.5 / telecom 57.24; standard submission by Sierra; reasoning effort: none; user simulator gpt-5.2; verified; rank #8 of 8 Official leaderboard | 61.6% |
tau3-Banking (tau-knowledge)
17 entries, in the board's order
| 1 | Qwen 3.8 Max (Qwen) tau3-Banking leaderboard (default tab on taubench.com), banking_knowledge pass^1; standard submission by Sierra; reasoning effort: xhigh; retrieval config: alltools; user simulator gpt-5.2; rank #1 of standard submissions Official leaderboard | 55.2% |
| 2 | Claude Opus 5 (Anthropic) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: max; retrieval config: alltools; user simulator gpt-5.2; rank #2 Official leaderboard | 48.7% |
| 3 | Grok 4.5 (xAI) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: high; retrieval config: alltools; user simulator gpt-5.2; rank #3 Official leaderboard | 47.9% |
| 4 | GPT-5.6-sol (OpenAI) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: xhigh; retrieval config: alltools; user simulator gpt-5.2; rank #4 Official leaderboard | 46.9% |
| 5 | GPT-5.5 (OpenAI) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: xhigh; retrieval config: alltools; user simulator gpt-5.2; rank #5 Official leaderboard | 44.6% |
| 6 | Muse Spark 1.1 (Meta) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: xhigh; retrieval config: alltools; user simulator gpt-5.2; rank #6 Official leaderboard | 40.5% |
| 7 | Claude Opus 4.7 (Anthropic) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: max; retrieval config: alltools; user simulator gpt-5.2; rank #7 Official leaderboard | 40.2% |
| 8 | Claude Opus 4.8 (Anthropic) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: max; retrieval config: alltools; user simulator gpt-5.2; rank #8 (tied 39.7 with Claude Fable 5) Official leaderboard | 39.7% |
| 9 | Claude Fable 5 (Anthropic) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: max; retrieval config: alltools; user simulator gpt-5.2; rank #9 (tied 39.7 with Claude Opus 4.8) Official leaderboard | 39.7% |
| 10 | GPT-5.4 (OpenAI) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: xhigh; retrieval config: alltools; user simulator gpt-5.2; rank #10 Official leaderboard | 39.4% |
| 11 | Kimi K3 (Moonshot AI) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: max; retrieval config: alltools; user simulator gpt-5.2; rank #11 (tied 37.1 with GLM-5.2) Official leaderboard | 37.1% |
| 12 | GLM-5.2 (Z.ai) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: xhigh; retrieval config: alltools; user simulator gpt-5.2; rank #12 (tied 37.1 with Kimi K3) Official leaderboard | 37.1% |
| 13 | GPT-5.2 (OpenAI) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: high; retrieval config: alltools; user simulator gpt-5.2; rank #13 Official leaderboard | 32.2% |
| 14 | Gemini 3 Flash (Google) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: high; retrieval config: terminal; user simulator gpt-5.2; rank #14 (tied 27.3 with Claude Opus 4.6) Official leaderboard | 27.3% |
| 15 | Claude Opus 4.6 (Anthropic) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: max; retrieval config: alltools; user simulator gpt-5.2; rank #15 (tied 27.3 with Gemini 3 Flash) Official leaderboard | 27.3% |
| 16 | Gemini 3.1 Pro Preview (Google) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: high; retrieval config: alltools; user simulator gpt-5.2; rank #16 Official leaderboard | 26.0% |
| 17 | Claude Sonnet 4.5 (Anthropic) tau3-Banking leaderboard, banking_knowledge pass^1; standard submission by Sierra; reasoning effort: enabled; retrieval config: terminal; user simulator gpt-5.2; rank #17 Official leaderboard | 25.3% |
About this board
- The current tau2-bench Overall board (retail + airline + telecom, standard submissions, non-legacy, as of 2026-10-17) lists only 8 entries, so a top 25 cannot be filled from it. The other 17 entries come from the tau3-Banking (tau-knowledge) board on the same official site. Each entry's board label says which board it is from, and the two scores are not comparable.
- Newer frontier models (Claude Opus 5, Claude Fable 5, Claude Opus 4.6/4.7/4.8, GPT-5.4/5.5/5.6-sol, Grok 4.5, Kimi K3, GLM-5.2, Qwen 3.8 Max, Muse Spark 1.1, Inkling, Gemini 3.1 Pro Preview) have only banking_knowledge results in the official submissions. They have no retail/airline/telecom tau2-bench Overall score on the official leaderboard.
- Hidden by default and not included: custom-scaffold submissions (Distyl ButtonAgent, banking 31.2%; RAFT-30B-A3B by NEU, retail only 82.46), legacy tau-bench v1 submissions (16 entries, not directly comparable because of task fixes), and the tau3-Voice board (24 voice submissions).
- Scores are from taubench.com/submissions/manifest.json and the per-submission submission.json files in github.com/sierra-research/tau2-bench at web/leaderboard/public/submissions (commit 96a3ef92329baf908ffe9dfe188cea34e8701d2a, 2026-10-15). The live manifest was fetched on 2026-10-17 and matched the repo. The board's Overall is computed as the mean of the three domain pass^1 values, and each score is formatted with toFixed(1), the same way the site renders it. Each entry's date is the submission date.
FAQ
What does tau-bench and tau2-bench measure?
tau-bench measures how well AI agents converse with users, call tools, retrieve knowledge, and follow policy across enterprise customer-service domains, in text and voice. This page covers 2 separate boards: tau2-bench (Overall) and tau3-Banking (tau-knowledge). Each is scored on its own terms, so compare rows only within the same board.
Which model leads tau-bench and tau2-bench?
As of October 17, 2026, Qwen3.5-397B-A17B (Alibaba Cloud) ranks first on the tau2-bench (Overall) board with 87.9% and Qwen 3.8 Max (Qwen) ranks first on the tau3-Banking (tau-knowledge) board with 55.2%. That is the board's own order, not a Developers Digest ranking.
Where do these tau-bench and tau2-bench scores come from?
All 25 entries come from the official leaderboard at taubench.com. Every entry links to its source, which was read on October 17, 2026, and lists the setting it was run under.
Related reading
- Kolibri Release Guide: Aleph Alpha's 78B Open ModelKolibri is an Apache 2.0 English-German MoE with 78B total and 3B active parameters and a 1M context. Benchmarks, the vLLM command, and who should run it.
- The Oracle Is Agreeing With ItselfSelf-improving test generation loops can grade against an oracle that agrees with itself. An audit shows the inflated gain and why independent checks count.
- Stop Means Stop: New Paper Finds Agent Approval Gates and Cancellation Leak in Six FrameworksA new arXiv paper probes six widely used open-source agent frameworks and finds the barrier semantics of approval gates, cancellation, and timeouts hold on none of them. A sibling branch can execute while the user is rejecting another one, and replay can double-execute. The fix is a verified external gate called SoundGate.
- Muse Glimmer 30B: Meta's Open-Weight Local Agent Model, Benchmarks, and Hardware RealityMuse Glimmer 30B is Meta's Apache 2.0 open-weight multimodal agent model that runs in 24GB. Benchmarks, hardware reality and what the numbers actually say.
- Weekly Highlights: Agents Became the Attack Surface, Open Weights Took the Agentic LeadThe Shai-Hulud worm hit 434 npm packages, three labs confirmed agents hit real targets, and Qwen 3.8 Max reached the agentic frontier with open weights.
- Sakana Fugu Ultra: The Model Router Making the Frontier Look Less ProprietarySakana Fugu Ultra is a learned orchestration layer routing work across expert models behind one API - a serious case for multi-model AI over one giant model.
Sources
Last verified . Each row links to the page its score was copied from.
Newsletter
Get new benchmark results as they land
One email per week. New model releases, the leaderboard moves that matter and what they mean for what you build.