The visible date on this Sierra post is October 21, 2026. The judgment first: τ-i18n extends the earlier τ-Voice benchmark to five languages, Spanish, Brazilian Portuguese, Hindi, Korean, and Mandarin. The airline, retail, and telecom tasks, policies, tools, and success criteria stay fixed, so the languages can be compared. Three voice agents and 2,700 primary calls are the scale the page gives. On task completion, Korean trails English by 22.6 points on average and Mandarin by 9.1. That gap is task completion. It is not the interaction score, and it is not the generation-quality score.
[1]
The three nearer languages drop less. Spanish, Hindi, and Portuguese trail English on task completion by 2.4, 3.8, and 6.4 points on average. Korean is the weakest language for all three agents. The page stresses that these calls run the same tasks. The drop therefore cannot be rewritten as a harder set of jobs. It is also not written as one model being unable to use a language at all. The average gap is a statement across the three agents, not a separate table for each one.
[1]Task completion says only whether the caller's goal was met. τ-i18n reports two further measures. Interaction asks whether the agent responds or yields at the right time, avoids interruptions and long monologues, can ignore backchannels and speech that is not directed at it, calls tools with the right arguments, and avoids errors it caused itself. Generation quality is what the caller actually hears: translationese, grammatical gender, register and honorifics, and whether the audio delivers the intended words, names, and numbers. The three are reported separately because an agent can finish the task and still fall short on the interaction or on the language.
[1]Interaction and task completion are not the same ranking. Korean and Mandarin calls have the most tool failures for every agent. Responsiveness is a common failure: GPT Live 1 and Gemini 3.8 Live miss a caller turn, reply slowly, or fail to yield in more than 80 percent of calls. For Grok Voice Think Fast 2.0 that figure is 16 percent. Selectivity is weak too: in 26 to 60 percent of calls, the agent responds more than three times to backchannels, vocal tics, or speech not directed at it. The range is what the page gives. The body does not split it into one row per agent.
[1]Generation quality splits the ranking again. GPT Live 1 leads both task completion, at 79.9 percent, and generation quality, at 63.4 percent. Grok Voice Think Fast 2.0 leads interaction at 52.6 percent and scores just 17.7 percent on generation quality. The page's conclusion is that the most responsive agent is not necessarily the most natural speaker. The figures 79.9, 63.4, 52.6, and 17.7 each sit on their own measure. They cannot be added into one total.
[1]The five languages are described as a starting point. The page says the codebase and the LanguageFactory behind the language packs are open sourced, so native speakers can adapt the packs: caller identities, matching database records, personas, voices, speech conventions, and evaluation rules, while the underlying service tasks stay the same. Korean and Mandarin callers, for example, typically give the family name first. An agent that puts the first spoken name into a given-name field can fail authentication even when it heard the name correctly. That is a failure mode the page names, not a counted authentication-failure rate. Further languages are an invitation, not a leaderboard already finished.
[1]