Quick Answer
Grok Voice 3.0 vs GPT-Live-1 vs Gemini 3.8 Live: Voice AI
The short answer
Grok Voice 3.0 now leads the published speech-to-speech benchmarks, Gemini 3.8 Live is the cheapest per minute, and GPT-Live-1 fits teams already building on OpenAI’s backend models. xAI shipped Grok Voice 3.0 on October 14, 2026 at an unchanged $0.08 per minute. Facts as of October 17, 2026.
Head to head
| Grok Voice 3.0 (xAI) | GPT-Live-1 (OpenAI) | Gemini 3.8 Live (Google) | |
|---|---|---|---|
| Released | October 14, 2026 | 2026 (API) | September 2026 |
| API price | $0.08/min ($4.80/h) flat | $0.05/min session, billed per second, + backend model and tool usage | $3/MTok audio in ( |
| AA Speech-to-Speech Index | 86.9 | 83.2 (Astra, medium) | 82.6 (extended thinking, high) |
| Big Bench Audio | 98.5% | 90.1% | 97.7% |
| Task success rate | 94.8% | 87.4% | 89.1% |
| Full Duplex Bench | 97.0% | 94.9% | 91.9% |
| Speech Agent Arena Elo | 1070 | 1048 | 990 |
Benchmark figures are from xAI’s launch post, which cites Artificial Analysis; prices are from each vendor’s pricing page.
What changed with Grok Voice 3.0
- Accuracy: +5.6 points on the AA Speech-to-Speech Index over Grok Voice 2.0 (81.3), and 71.6% on τ-Voice, a support-call benchmark with accents, noise and interruptions — comparable with GPT-Live-1 at about 40% lower latency (xAI).
- Enterprise result: ServiceNow’s EVA-Bench scored it 74%, the highest accuracy across 200+ enterprise scenarios.
- Production scale: xAI says Starlink call centres use it for 20,000 calls a day across languages, with roughly 70% resolved without a human.
- No migration: same price as 2.0;
grok-voice-latestagents moved automatically.
Cost per minute, honestly
The three are billed differently, so headline rates mislead:
- Grok Voice 3.0 is flat: $0.08 a minute covers the conversation; text inputs and tools such as web search and file search are billed separately.
- GPT-Live-1’s $0.05 covers the voice layer only. When it calls a backend reasoning model or tools, those tokens are charged at their own rates — a long tool-using call can cost more than Grok’s flat rate.
- Gemini 3.8 Live is per token. At Google’s conversion a minute of input plus a minute of output is about $0.023, but thinking tokens count as output, so heavy reasoning raises it.
Compare cost per resolved call on 50 recorded scenarios, not per-minute list prices.
Which to pick
- Customer support or sales calls where accuracy matters most: Grok Voice 3.0.
- Agent already runs on OpenAI models, tools and Agents SDK: GPT-Live-1.
- High-volume, cost-sensitive, or needs Google Search grounding: Gemini 3.8 Live.
- Need a full platform (telephony, testing, analytics) rather than a model: build on one of these through a voice-agent platform — see Retell vs Vapi vs Bland.
Related: best speech-to-speech AI models, Gemini 3.8 Live vs GPT-Live-1, and text-model rates in current API prices.
Last verified: October 17, 2026.