Benchmarks / tau2-bench

Reported by tau2-bench

tau2-bench

Share of banking questions answered from a policy knowledge base tasks completed within policy in all four trials.

Data older than 90 days: last updated 3 Aug 2026

Results dated
24 Feb 2026 to 2 Mar 2026
Models
3
Unit
% of tasks
Licence
MIT License

tau2-bench in brief

What it measures. tau2-bench is a third-party benchmark, shown here with credit. Measured here: Share of banking questions answered from a policy knowledge base tasks completed within policy in all four trials.

Latest results. , 3 models ranked.

Top 3 on Consistency (pass^4) (% of tasks, higher is better):

  1. Claude Sonnet 4.5 (reasoning on) (Anthropic): 10.3%
  2. Gemini 3 Flash Preview (high reasoning) (Google): 7.2%
  3. Gemini 3 Pro Preview (high reasoning) (Google): 4.1%

How to cite. tau2-bench, via Spring Prompt, results dated 2 March 2026. https://springprompt.com/benchmarks/tau2-bench These results belong to tau2-bench, not to us. Licence: MIT License. Credit tau2-bench first.

Consistency (pass^4): Gemini 3 Pro Preview

3 results · % of tasks, higher is better. Choose a model to highlight it.Clear highlight

  1. 1 Claude Sonnet 4.5 (reasoning on)Anthropic 10.3%
  2. 2 Gemini 3 Flash Preview (high reasoning)Google 7.2%
  3. 3 Gemini 3 Pro Preview (high reasoning)Google 4.1%

Full results

tau2 v1.0.1, banking knowledge, terminal retrieval: consistency (pass^4), % of tasks, higher is better
#ModelConsistency (pass^4)
% of tasks, higher is better
Price
$ per million tokens, in / out
1 Claude Sonnet 4.5 (reasoning on)Anthropic
10.3%
$3 / $15
2 Gemini 3 Flash Preview (high reasoning)Google
7.2%
$0.50 / $3
3 Gemini 3 Pro Preview (high reasoning)Google
4.1%
–

Swipe the table sideways for more columns.

Source: tau2-bench, via Spring Prompt, springprompt.com/benchmarks/tau2-bench. Test your own prompt on these models →

Ranks follow the score as shown, so equal numbers share a rank. Results as published by tau2-bench; we do not re-run them.

Progress over time

Consistency (pass^4), plotted at each model's release date, higher is better. Each model counts once, at its best setting. Scores are today's, not launch day's. Hover or tap a mark for its model.

Best score so far

Every model at its release date. The line moves each time a new release beats every earlier one, and each new best is named where there is room.

2%4%6%8%10%12% Oct 20252026AprJulOct Claude Sonnet 4.5

Best score by lab

Each lab's best so far, stepping whenever one of its releases beats its own earlier models.

AnthropicGoogle

2%4%6%8%10%12% Oct 20252026AprJulOct AnthropicGoogle

What it measures

Share of banking questions answered from a policy knowledge base tasks completed within policy in all four trials.

What it does not measure

Not your policies or customers: the tasks and policies are tau2's own, and the customer is simulated by gpt-5.2.

2 results from tau2-bench not ranked here · show why

We rank a result only when we can tie it to a specific model you can use. These are left out:

  • A custom agent harness, not the standard one: 2

tau2-bench by Sierra Research and contributors, MIT License.