October 15, 2026 | AI | By Mahesh
xAI released Grok Voice 3.0 on October 14, 2026, a speech-to-speech model sold through its Voice Agent API at $0.08 per minute, the same price as Grok Voice 2.0 [1][2]. It scores 86.9 on the Artificial Analysis Speech-to-Speech Index, first place among the models xAI compared, ahead of OpenAI's GPT-Live-1 at 83.2 and Google's Gemini 3.8 Live at 82.6 [1]. xAI also says the model resolves more support calls than rivals at lower cost. Most of the evidence comes from xAI itself, and on one support-call test the model trails OpenAI.
- Grok Voice 3.0 is xAI's new speech-to-speech model for voice agents, and it replaces Grok Voice 2.0 automatically for apps using the latest alias.
- It costs $0.08 per minute, unchanged from version 2.0, with concurrency limits tied to how much you have spent.
- It ranks first on one independent index but second on a support-call test, and most deployment figures are xAI's own.
What is Grok Voice 3.0 and what is new
Grok Voice 3.0 is xAI's latest speech-to-speech model, which listens to spoken audio and answers in speech through a single live connection [1][2]. xAI calls it "the world's most intelligent speech-to-speech model" and claims sub-second response times, a record for resolving customer support calls and a lower running cost than other frontier models [1]. Grok Voice 2.0 is now deprecated [4].
xAI points to two deployments. Tesla uses the model for millions of in-car conversations a day, including calendar and email checks and Grok Bot's voice mode. Starlink's call centers use it to answer about 20,000 calls a day in several languages, where xAI says roughly 70% are resolved without a human and one in four sales calls ends in a purchase before the caller hangs up [1][3]. Those figures are xAI's claims and no outside audit accompanies them.
"Grok Voice 3.0 sets a new Pareto frontier on EVA-Bench"Nitin Aggarwal, Senior Director of Product Management, Agentic Evals and RAI, ServiceNow, in xAI's announcement [1]
The model is a different product from the text models Depth Grid has covered recently. Meta's Muse Wonder accepts audio input but answers in text only, as covered in Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks, so it does not compete for live phone calls.
Under the hood: API, voices and migration
xAI has not disclosed a parameter count or an architecture beyond "speech-to-speech." Developers connect over a WebSocket at wss://api.x.ai/v1/realtime, send a session.update message with instructions, a voice and tools, and then stream audio [2][4]. Browser and mobile apps should use ephemeral tokens instead of API keys. The model id is grok-voice-3.0, and the grok-voice-latest alias now points to it, so agents using the alias moved to 3.0 with no code changes. One guide recommends pinning the versioned id in production [4].
The API supports file_search, web_search, x_search, MCP and custom function tools, plus a reasoning.effort setting that defaults to high and can be turned off [4]. There are 28 built-in voices with eve as the default [2]. Language support is unclear. The docs list 20 languages with native accents and the landing page says more than 25, and the model detects the caller's language on its own [4].
Phone calls run over SIP through a carrier trunk you already own, with setup guides for Twilio and Telnyx. Phone numbers cannot be provisioned through the API [4]. Version 3.0 runs in three regions, us-east-1, eu-west-1 and us-saltlake-2, where version 2.0 ran only in us-east-1 [4].
Benchmarks: first on one index, second on another
The table uses Artificial Analysis data as of October 13, 2026, as published by xAI, plus the τ-Voice support-call test [1][3]. Artificial Analysis owns the Speech-to-Speech Index, but the article it appears in does not say who ran xAI's other tests [5].
| Benchmark | Grok Voice 3.0 | Grok Voice 2.0 | GPT-Live-1 (Astra, Medium) | Gemini 3.8 Live |
|---|---|---|---|---|
| Speech-to-Speech Index | 86.9 | 81.3 | 83.2 | 82.6 |
| Big Bench Audio (speech reasoning) | 98.5% | 97.2% | 90.1% | 97.7% |
| Speech Agent Arena Elo (human preference) | 1,070 | 1,011 | 1,048 | 990 |
| Task success rate | 94.8% | 94.6% | 87.4% | 89.1% |
| Full Duplex Bench (turn-taking) | 97.0% | 95.1% | 94.9% | 91.9% |
| τ-Voice (simulated support calls) | 71.6% | 56.5% | 74.5% | 68.6% |
The gain over Grok Voice 2.0 is clear. The index rises from 81.3 to 86.9 and τ-Voice from 56.5% to 71.6% [3]. xAI also reports 74% accuracy on ServiceNow's EVA-Bench, the highest score there, and a cost of $0.025 per Big Bench Audio task against $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1 [1].
Two details temper the headline. On τ-Voice the model scores 71.6% against 74.5% for GPT-Live-1, which xAI's text calls "comparable." The xAI landing page claims first place on the τ-Voice leaderboard, but Artificial Analysis ranks the model second there and third on human-preference Elo [3][4].
Pricing, availability and how to access it
Grok Voice 3.0 costs $0.08 per minute, or $4.80 per hour, in the Voice Agent API, unchanged from version 2.0 [1][2]. Tool calls are billed separately at $2.50 per 1,000 for file search and $5 per 1,000 for web search. For scale, 36,000 minutes a month comes to $2,880 in voice time before tool and carrier fees [4]. Separate speech-to-text and text-to-speech APIs are priced at $0.20 per hour streaming and $15 per million characters [2].
Concurrent sessions depend on lifetime spend since January 1, 2026. Tier 0 allows 10 sessions, Tier 1 ($50) allows 20, Tier 2 ($250) allows 50, Tier 3 ($1,000) allows 100 and Tier 4 ($5,000) allows 200, with higher limits by request [4]. A new account therefore cannot load-test 50 simultaneous calls without spending enough to reach Tier 2.
To try it, create a key in the xAI Console and use the Playground or the no-code Voice Agent Builder. The xAI announcement and Voice Agent API docs have the setup details [1][2]. There are no open weights.
Limitations, trade-offs and alternatives
Latency claims do not line up. xAI's documentation lists time to first audio at 0.81 seconds. One summary says that is slower than the 0.70 seconds of version 2.0, while another says it fell from 1.34 seconds, and xAI's page claims 40% lower latency than GPT-Live-1 without a figure [1][4][5]. Test it on your own calls before relying on any of these numbers.
Migration needs care. xAI publishes no verbosity metric, and its notes recommend shorter system prompts. Some OpenAI Realtime events are named differently or not emitted, so an app built for that API needs testing [4]. The eesel guide cited for these details is published by a company that sells a competing helpdesk AI product, so weigh its comparisons accordingly.
Alternatives are OpenAI's GPT-Live-1 and GPT-Realtime-2.1 and Google's Gemini 3.8 Live, which the benchmark table above covers. A voice agent also needs a text model for planning and tool use, and Anthropic's small model is one cheap option, priced in Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks.
Sources
- xAI. "Introducing Grok Voice 3.0," October 14, 2026. x.ai
- xAI. "Voice Agent API," developer documentation. docs.x.ai
- Unite.AI. "xAI Introduces Grok Voice 3.0 Speech-to-Speech Model for Voice Agents." unite.ai
- eesel AI. "Grok Voice 3.0: what it is, what changed, and how to use it." eesel.ai
- AI Weekly. "xAI Grok Voice 3.0 Ships at $0.08/Min, Tops Index at 86.9." aiweekly.co
Read More on Depth Grid
- Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks
- Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks
- Tech News Today, Oct 14: Character.AI, OpenAI, Meta, Bumble
- Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
Article by Mahesh | Depth Grid
Discussion (0)
Comments are disabled for this post.