Grok Voice 3.0 has taken first place on Artificial Analysis’s current Speech-to-Speech Index with a score of 86.9. The result combines four different capabilities: answering spoken reasoning questions, completing tool-using support calls, winning blind human-preference comparisons and reaching the correct task outcome. Its predecessor scored 81.3.[1][2]
The composite lead does not mean it wins every constituent test. On τ-Voice—the benchmark most directly built around noisy simulated customer-service calls—Grok Voice 3.0 scored 71.6%, behind GPT‑Live‑1 at 74.5% in the cited comparison. The new model is better than Grok Voice 2.0 by 15.1 percentage points on that test, but “best overall” and “most reliable on this support workflow” remain different claims.[1][2]
A pleasant call and a completed task are separate outputs
A voice agent can respond quickly, avoid awkward pauses and sound natural while still changing the wrong booking or failing to issue a credit. That is why the benchmark suite separates conversation quality from business state. Full Duplex Bench measures behaviours such as interruption handling and turn-taking. τ-Voice evaluates whether a tool-using conversation leaves the simulated company database in the one correct end state.
What the headline numbers measure
| Measure | Grok Voice 3.0 | Meaning |
|---|---|---|
| Speech-to-Speech Index | 86.9 | Equal-weight composite of four components in the current v2.0 methodology.Source 1Source 2 |
| Big Bench Audio | 98.5% | Spoken reasoning accuracy under the benchmark’s grading protocol.Source 1Source 2 |
| Speech Agent Arena | Elo 1070 | Blind preference result summarized as an Elo rating.Source 1Source 2 |
| Task Success Rate | 94.8% | Share of eligible conversations ending with the correct task-completing tool call.Source 1Source 2 |
| τ-Voice | 71.6% | Simulated support scenarios judged by the final database state.Source 1Source 2 |
| Full Duplex Bench | 97.0% | Conversational dynamics reported separately; no longer part of the v2.0 composite.Source 1Source 2 |
The composite changed in August
Artificial Analysis’s version 2.0 index gives 25% weight to each of speech reasoning, τ-Voice, Arena preference and Task Success Rate. It replaced Full Duplex Bench inside the composite, although the conversational score is still published. A reader comparing an older index article with today’s number can therefore be comparing different formulas even when the model name is unchanged.[2]
Current comparison cited by xAI
| Model | Composite | τ-Voice | Task success | Full duplex |
|---|---|---|---|---|
| Grok Voice 3.0 | 86.9 | 71.6% | 94.8% | 97.0%Source 1 |
| Grok Voice 2.0 | 81.3 | 56.5% | 94.6% | 95.1%Source 1 |
| GPT‑Live‑1, Astra medium | 83.2 | 74.5% | 87.4% | 94.9%Source 1Source 2 |
| Gemini 3.8 Live, extended thinking high | 82.6 | 68.6% | 89.1% | 91.9%Source 1Source 2 |
The most important improvement is not the 5.6-point composite increase by itself. τ-Voice rose from 56.5% to 71.6%, a 15.1-point gain. Because the current composite weights that component at one quarter, the score change in that test contributes about 3.8 composite points if all else is held fixed. That is our arithmetic on the published figures, not evidence about which training change produced the improvement.[1][2]
Latency creates a real operational trade-off
xAI says Grok’s τ-Voice score is comparable with GPT‑Live‑1 while running at 40% lower latency. Faster first audio can make a call feel more competent and reduce interruption overlap. It cannot compensate for an incorrect account change. A deployment needs a threshold for both: how quickly the system responds and how often it completes each high-consequence task correctly.[1]
API economics before business overhead
| Measure | Value | Context |
|---|---|---|
| Audio price | $0.08 per minute | Equivalent to $4.80 per hour of billed audio under the published tariff.Source 3 |
| Text event | $0.004 per qualifying text input | Text-only conversation.item.create events are charged separately.Source 3 |
| Default turn detection | Session duration is billed | Push-to-talk is billed for audio sent and received instead.Source 3 |
| Model alias | grok-voice-latest | xAI says agents using the alias were automatically moved to version 3.0.Source 1 |
| Region | us-east-1 | The documentation lists the speech-to-speech region for the API.Source 3 |
At the tariff rate, a ten-minute audio session costs $0.80 before text events and any other connected services. That is not the cost of resolving a case. A failed call may create a second call, human escalation, refund, compliance review or customer loss. The economically relevant denominator is cost per correctly completed outcome, segmented by task and consequence.[3]
Production claims need a second ledger
xAI says Starlink uses Grok Voice for 20,000 calls a day, resolves about 70% without a human and converts one in four sales calls. Those are company-reported operational figures from a related business, not an independent audit or proof that another company will reproduce them. The model also operates with Starlink’s tools, policies and data, which are part of the system result.[1]
Grok Voice 3.0 is a substantial release because it improves both the composite and the support benchmark while preserving the published per-minute price. Its numbers also demonstrate why a single voice ranking is insufficient. A customer hears one conversation. The business inherits the database state left behind.
Sources & further reading
- Introducing Grok Voice 3.0 SpaceXAI · 14 Oct 2026 · Accessed 16 Oct 2026
- Speech-to-Speech Benchmarking Methodology Artificial Analysis · Accessed 16 Oct 2026
- Speech to Speech pricing and rate limits SpaceXAI Docs · Accessed 16 Oct 2026