Free🇭🇰🇨🇳
Chinese evaluation suite with 13,948 multiple-choice questions in 52 disciplines and four difficulty levels for foundation models.
τ-bench (tau-bench) overview and key facts (October 2026)
Sierra's benchmark for AI agents that talk to users, call tools and follow policy in retail, airline, telecom and banking.
Free
What is τ-bench (tau-bench)?
τ-bench tests whether AI agents can reliably handle customer service conversations: talking with users, calling tools, retrieving knowledge and following company policy, in text and voice.
The benchmark covers retail, airline, telecom and banking, plus a voice domain with interruptions and accents. Results are checked against the final state of a database, so an agent must actually complete the task. τ³-bench, the 2026 release, added a knowledge-retrieval banking domain with configurable RAG pipelines, full-duplex voice evaluation and more than 75 task fixes. Sierra maintains the benchmark, the MIT-licensed code on GitHub and the leaderboard at taubench.com.
τ-bench (tau-bench) AI features and use cases
- Tests service agents before launchCustomer service teams can run their planned agent through realistic conversations in domains close to their own. The pass^k metric shows whether the agent succeeds consistently across repeated attempts. Consistency matters in service.
- Checks policy complianceTasks require the agent to follow rules such as refund or change policies while using tools. Failures show where an agent would break company rules with real customers. These cases can then be addressed before launch.
- Evaluates voice agentsThe full-duplex voice domain tests real-time conversations with interruptions and accents. Contact center teams can compare voice models before deploying them on phone lines. Voice results are shown on the same leaderboard as text results.
What sets τ-bench (tau-bench) apart
- Scores agents on the resulting database state and on consistency across repeated runs.
- Includes voice and knowledge-retrieval domains in the 2026 release.
Where τ-bench (tau-bench) is strongest
Available internationally
τ-bench is maintained by Sierra and used worldwide through its public leaderboard at taubench.com.
Questions to ask before you buy τ-bench (tau-bench)
Sierra sells customer service agents itself, which is worth knowing when reading the leaderboard. The domains are simulated and in English, so test your agent on your own policies and languages as well. Scores on banking_knowledge changed with version 1.0.1, so compare only results from the same version.
τ-bench (tau-bench) reviews
No reviews yet. Be the first to share how your team uses it.
Every review shows its verification level. We read each one before it goes live. How we check reviews
Questions about τ-bench (tau-bench)
Which domains does τ-bench cover?
Retail, airline, telecom and banking, plus a voice domain with real-time conversations, interruptions and accents.
What does pass^k mean?
It measures whether an agent completes the same task successfully across k repeated attempts, which shows reliability rather than a lucky single run.
Is the code open?
Yes. The tau2-bench repository is published by Sierra under the MIT license, with instructions for leaderboard submissions.
τ-bench (tau-bench) update historyno updates yet
No changes recorded for τ-bench (tau-bench) yet.
Changes the vendor made and confirmed user reports, each checked against the vendor's own pages. We review every tool monthly.
Alternatives to τ-bench (tau-bench)
Tools in the same category (Domain benchmarks) with the most overlapping markets.
Free🇺🇸
Collaborative benchmark of 162 English legal reasoning tasks from 40 contributors, covering contracts, statutes and court opinions.
Free🇨🇳
Open-source LLM evaluation platform from Shanghai with CompassRank leaderboards, the CompassHub benchmark catalog and VLMEvalKit.
Free🇭🇰
Benchmark of 369 real computer tasks on Ubuntu, Windows and macOS that tests multimodal agents operating desktop and web apps.