# τ³-Banking

Text agents resolve banking customer-service tasks by retrieving from a knowledge base of about 700 documents. Published as τ-Knowledge; the banking_knowledge domain of τ³-bench.

## Facts

- Maker: Sierra (United States)
- Year: 2026
- Use case: Finance knowledge
- Size: Not stated
- Public split: Code and tasks in the open tau2-bench repository
- Grading: pass^k: the share of tasks solved in all of k trials, with an LLM user simulator. Headline is pass^1.
- Licence: MIT (code)
- Submission: Submissions to the τ-bench leaderboard
- Products allowed: Not stated
- What it does not cover: Customer service only: no investment analysis, trading or filings.
- Last checked: 2 Nov 2026
- Links: paper https://arxiv.org/abs/2603.04370, site https://taubench.com, code https://github.com/sierra-research/tau2-bench, leaderboard https://taubench.com/leaderboard/?benchmark=knowledge
- Directory page: https://benchmarks.finance/benchmarks/tau3-banking/
- Leaderboard: https://benchmarks.finance/leaderboards/tau3-banking/

## Results

### τ³-Banking: pass^1 % (higher is better, 5 models)

| Rank | Model | Provider | Score | Metric | Setting | Reported by | Source | URL | Date |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | Qwen 3.8 Max | Qwen | 55.2 | pass^1 % | all tools retrieval, gpt-5.2 user simulator | Official board | τ-bench leaderboard (Sierra) | https://taubench.com/leaderboard/?benchmark=knowledge | 4 Aug 2026 (checked 2 Nov 2026) |
| 2 | Claude Opus 5 | Anthropic | 48.7 | pass^1 % | all tools retrieval, gpt-5.2 user simulator | Official board | τ-bench leaderboard (Sierra) | https://taubench.com/leaderboard/?benchmark=knowledge | 4 Aug 2026 (checked 2 Nov 2026) |
| 3 | Grok 4.5 | xAI | 47.9 | pass^1 % | all tools retrieval, gpt-5.2 user simulator | Official board | τ-bench leaderboard (Sierra) | https://taubench.com/leaderboard/?benchmark=knowledge | 4 Aug 2026 (checked 2 Nov 2026) |
| 4 | GPT-5.6 Sol | OpenAI | 46.9 | pass^1 % | all tools retrieval, gpt-5.2 user simulator | Official board | τ-bench leaderboard (Sierra) | https://taubench.com/leaderboard/?benchmark=knowledge | 4 Aug 2026 (checked 2 Nov 2026) |
| 5 | GPT-5.5 | OpenAI | 44.6 | pass^1 % | all tools retrieval, gpt-5.2 user simulator | Official board | τ-bench leaderboard (Sierra) | https://taubench.com/leaderboard/?benchmark=knowledge | 5 May 2026 (checked 2 Nov 2026) |


Source: https://benchmarks.finance/benchmarks/tau3-banking/

Created and maintained by Moneda.