First overall, second on the support test

xAI's launch post calls Grok Voice 3.0 the world's most intelligent speech-to-speech model and says it sets a new record for resolving customer support calls. The first half of that sentence has a table behind it. The second half depends on which table you choose.

On Artificial Analysis's composite Speech-to-Speech Index, the lead is real: 86.9, clear of GPT-Live-1 (Astra, Medium) at 83.2 and Gemini 3.8 Live at 82.6, and a solid jump from Grok Voice 2.0's 81.3. Speech reasoning on Big Bench Audio rises to 98.5%, and the Full Duplex score, which tests pauses, turn-taking and interruptions, reaches 97.0%.

Grok Voice 3.0 leads Artificial Analysis's composite Speech-to-Speech Index at 86.9. Its support-call record claim rests on ServiceNow's EVA-Bench, where xAI reports a leading accuracy score of approximately 74%. On Artificial Analysis's separate τ-Voice benchmark, which simulates customer-support interactions with varied accents, noise and interruptions, Grok scores 71.6%, behind GPT-Live-1 (Astra, Medium) at 74.5%. The results are not contradictory: they measure different evaluations under different conditions.

The quieter numbers

On Artificial Analysis's Speech Agent Arena task-success measure, Grok Voice rises only from 94.6% to 94.8%, despite a 5.6-point increase in the composite index. That suggests improvements elsewhere in the benchmark mix, but it does not mean customer-support task completion barely improved: τ-Voice rises substantially, from 56.5% to 71.6%. Average time to first audio is 0.81 seconds, faster than GPT-Live-1 (Astra, Medium) at 1.34 seconds, but slower than Grok Voice 2.0's 0.70 seconds. The 0.70-second figure is the one xAI published for Voice 2.0, as read from the announcement charts by iClarified. Still fast, still ahead of the named rival, and slightly slower than the model it replaces.

What it costs, and what changed without asking

The price did not move. Speech-to-speech stays at $0.08 per minute, $4.80 per hour, with a separate $0.004 charge per qualifying text input on xAI's pricing page.

xAI reports an average Big Bench Audio cost of $0.025 per task for Grok Voice 3.0, versus $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1. Those are costs for a particular benchmark workload, not estimates of an average customer-support call. Actual API spending depends on billable session duration, text-input events, retries and application design. The billing detail matters too: sessions using default server-side voice-activity detection are billed by session duration, while push-to-talk sessions are billed for audio sent and received; the $0.004 text-input charge applies to qualifying text events, with documented exceptions, per xAI's documentation.

xAI says Grok Voice powers millions of daily conversations across its deployments, including Tesla vehicles, and reports that Starlink handles approximately 20,000 support calls a day using Grok Voice, with roughly 70% resolved without human intervention. These are company-reported operating figures; xAI has not supplied independently audited call volumes, resolution definitions or comparable human-agent baselines in the launch announcement.

xAI says grok-voice-latest now resolves to Grok Voice 3.0 automatically, with no API changes required, while the explicit grok-voice-2.0 model is deprecated. Applications using the moving alias should verify their current model and rerun regression tests for call completion, interruptions, tool use, latency and cost. Applications already pinned to a specific model version are not automatically covered by the alias change. Pinning a version, where the API allows it, is the difference between evaluating a new model and discovering it in production.

Sources

  1. [1] xAI — "Introducing Grok Voice 3.0", launch post, 14 October 2026 (benchmark table, price, deployment figures, alias migration)Read source
  2. [2] xAI — API pricing documentation (checked 14 October 2026): speech-to-speech at $0.08/min plus $0.004 per text input, session billing rules, Grok Voice 2.0 listed as deprecatedRead source
  3. [3] xAI — Developer release notes (grok-voice-latest alias migration and model deprecations)Read source
  4. [4] Artificial Analysis — Speech-to-Speech Index and τ-Voice benchmark (testing dated 13 October 2026, cited by xAI)Read source
  5. [5] ServiceNow AI — EVA-Bench, the enterprise voice-agent benchmark behind xAI's support-call record claimRead source
  6. [6] iClarified — "Grok Voice 3.0 Launches With Improved Call Handling and Sub-Second Latency", 14 October 2026 (time-to-first-audio figures: 0.81s vs 1.34s; Voice 2.0 at 0.70s)Read source
  7. [7] RuntimeWire — "SpaceXAI ships Grok Voice 3.0, ranked No. 1 in speech-to-speech tests", 14 October 2026Read source
  8. [8] Unite.AI — "xAI Introduces Grok Voice 3.0 Speech-to-Speech Model for Voice Agents", 14 October 2026Read source