What happened
The public τ-bench leaderboard lists Ava Zero from Avartha in first place on its τ³-Voice Multilingual board. The displayed score is 86.0% task completion. The board covers all languages and three domains, according to the leaderboard page.
Two other systems are shown directly behind it: Phonic’s phonic-1 at 83.2%, and OpenAI’s gpt-live-1 at 79.9%. On this board, the ordering is therefore Ava Zero first, phonic-1 second, and gpt-live-1 third. Ava Zero’s displayed lead over phonic-1 is 2.8 percentage points, and its lead over gpt-live-1 is 6.1 points. The source presents the result as a leaderboard state rather than as a product launch, pricing announcement, or independent deployment study.
τ-bench describes itself as a benchmark for whether AI agents can reliably complete real-world tasks. Its stated evaluation surface is broader than producing a plausible reply: agents converse with users, call tools, retrieve knowledge, and follow policy across enterprise domains. The site says it evaluates text and voice settings. Its homepage also separates multiple benchmark surfaces: τ³-Banking for text and knowledge retrieval, τ³-Voice for real-time voice, and τ²-bench for text tasks in retail, airline, and telecom.
For the voice track, τ-bench describes real-time voice as full-duplex conversation involving interruptions, accents, and background noise, while retaining the benchmark’s task-oriented rigor. The site names retail, airline, telecom, and banking among voice-mode domains. The multilingual τ³-Voice leaderboard itself is labelled as covering three domains; the page does not, in the supplied material, specify which three feed this particular aggregate. Nor does it say how individual languages contribute to the reported all-languages result.
The score label is task completion, which makes the result distinct from a standalone speech-recognition, synthesis, latency, or subjective conversational-quality comparison. It indicates that the benchmark observed successful completion under its own task setup. A high aggregate score can therefore be relevant to teams building systems that must move from spoken conversation to an action or policy-compliant resolution. The source does not provide a per-language breakdown, per-domain breakdown, cost figure, latency figure, transcript set, model configuration, or error analysis for Ava Zero on the leaderboard page.
τ-bench has evolved from its original 2024 tool-agent-user interaction benchmark. The site says the original version scored agents against verifiable database outcomes with its Pass^1 reliability metric. Later updates added telecom tasks in which users can act on the world too, audited and fixed more than 50 airline and retail tasks, and added knowledge-intensive banking tasks using a realistic knowledge base of roughly 700 documents. The site dates the task-audit update to February 2026 and the knowledge and real-time-voice additions to March 2026. Those details explain the benchmark’s emphasis on task completion, but they are not evidence that every production voice deployment will see the same ranking.
What the archive adds
The Voxtral 3 field note covered a voice workflow that packages speech models with a real-time agent SDK, making orchestration and deployment part of the product boundary. This is a new benchmark fact in that broader voice-agent thread: it reports comparative task completion, not a new SDK or workflow release.
Trend signal
No supplied pattern genuinely fits this leaderboard result closely enough to support a pattern claim.
Why it matters
A team selecting a voice-agent stack should care about whether an agent completes the business task after conversation, tool use, knowledge retrieval, and policy constraints—not merely whether its audio sounds natural. Ava Zero’s leading 86.0% result makes it a candidate for an evaluation shortlist alongside phonic-1 and gpt-live-1.
That is not a procurement verdict. The gap from the next listed system is 2.8 percentage points, while the supplied source does not explain variance, test volume, operating conditions, or how the systems were configured. Still, the leaderboard gives builders a concrete evaluation criterion: test the entire voice-agent loop against verifiable outcomes, including interruptions and tool actions, rather than treating the language or speech model as the sole product decision.
Limits and uncertainty
The supplied public page establishes a leaderboard placement and displayed task-completion scores. It does not establish how Ava Zero is built, what models it uses, whether its configuration is reproducible, or the cost and latency required to achieve the result. It also does not supply scores by language, domain, task, or failure type, so it cannot show where the system is strong or weak.
Benchmark performance is not equivalent to production reliability. A team’s tools, policies, customer accents, knowledge base, permissions, escalation paths, and success criteria may differ substantially from τ-bench. The source also provides no indication of whether the ranking will remain stable after future submissions or benchmark revisions.
Skill to build now
Create a voice-agent scorecard around completed work. Choose a small set of representative calls, define an externally verifiable success condition for each, and record whether the agent handled the conversation, tool action, policy constraint, and recovery path. Track completion separately from latency, speech quality, and cost. That produces a comparison closer to the τ-bench framing than an anecdotal demo review.
Reader takeaway
Ava Zero currently leads τ-bench’s multilingual voice-agent leaderboard at 86.0% task completion, ahead of phonic-1 and gpt-live-1. Treat that as a useful shortlist signal, then validate the full voice-and-tools workflow against your own tasks before choosing a stack.
Open question
Will future τ³-Voice results publish the language, domain, configuration, cost, and failure breakdowns needed to turn an aggregate leaderboard placement into a deployment decision?