sierra-research/tau2-bench released v2.0.0. τ-bench is a benchmark that covers how tools, agents and users interact with one another in real domains. What this version adds is text evaluations and voice evaluations in six languages, plus a fixed benchmark configuration. The release date is October 20, 2026, and the details are summarized on the GitHub v2.0.0 release tag page.
What this benchmark measures is the interaction between agents that use tools and users within the same domain. It bundles text evaluations and voice evaluations together across six languages, and it also includes a fixed benchmark configuration that pins the evaluation conditions and procedure to a single standard. The fixed configuration is a way of packaging the evaluation setup so it can be reused repeatedly under the same conditions. The main point of this version is that the number of evaluation languages has grown and a fixed configuration now exists. The repository has 2275 stars.
However, the release notes state explicitly that 2.0 scores cannot be compared directly with 1.x. Version 2.0 adds text and voice evaluations in six languages and a fixed configuration, and its scoring basis is not the same as 1.x. The notes advise using v1.0.1 if you want to compare against the previous baseline. The two numbers cannot be placed side by side in a single line. The published release information contains no per-language scores, no per-model performance figures, and no material bridging the two baselines. It is safer to keep the comparison baseline at v1.0.1 and to look at the 2.0 results separately.