Grok Voice 3.0 tops the Speech-to-Speech Index at $0.08/min, but the bill counts silence
On October 14, 2026, xAI released Grok Voice 3.0 in its Speech to Speech API, the product it launched in December 2025 as the Voice Agent API. Agents pointed at grok-voice-latest were switched over automatically. Price is unchanged from Grok Voice 2.0 at $0.08 per minute. The headline claim is first place on Artificial Analysis' Speech-to-Speech Index with a score of 86.9, and I checked it on the leaderboard itself.
Teams running phone agents care less about the leaderboard than about two other things. One is the flat per-minute price, which makes call costs predictable in a way token-metered realtime APIs don't. The other is an alias upgrade that changed your production model without a deploy. If you're on grok-voice-latest, you've been running 3.0 for ten days whether you tested it or not.
What actually shipped
From xAI's announcement and docs:
| Item | Detail |
|---|---|
| Model | Grok Voice 3.0, served through wss://api.x.ai/v1/realtime?model=grok-voice-latest. No pinned model ID is given in the announcement. |
| Price | $0.08/min ($4.80/hr). With the default server_vad turn detection, you're billed for session duration. Push-to-talk sessions are billed only for audio sent and received. Text input costs $0.004 per conversation.item.create event. function_call_output items are not billed. |
| Limits | 10 concurrent sessions per team, 120-minute maximum session, a single listed region (us-east-1). |
| Protocol | The docs say it's compatible with the OpenAI Realtime API. Event names don't fully match (for example, the transcription delta event is named differently). |
| Telephony | SIP, WebSocket or LiveKit, with native G.711 μ-law/A-law at 8 kHz, so no transcoding. |
| Tools | Custom function calling, plus server-side web search, X search, collections search and remote MCP. |
| Languages | The docs say "20+", and the product page says "25+". |
The billing detail in the price row is the one that matters: on server_vad, the meter runs on session duration. For context, the December 2025 launch post priced the API at a flat $0.05/min. It has since gone up 60%. I couldn't find when the change happened.
Numbers that matter
The Artificial Analysis Speech-to-Speech Index combines four components: Big Bench Audio (speech reasoning), τ-Voice (agentic customer-service tasks), a blind human-preference Arena, and Task Success Rate (the correct final tool call). Top of the leaderboard as I read it:
| Model | Index | Avg time to first audio | AA cost / hr (normalized) |
|---|---|---|---|
| Grok Voice 3.0 | 86.9 | 0.81 s | $4.80 |
| GPT-Live-1 (Astra, medium) | 83.2 | 1.34 s | $5.83 |
| Gemini 3.8 Live Ext. Thinking (High) | 82.6 | 1.35 s | $3.50 |
| Grok Voice 2.0 | 81.3 | 0.70 s | $4.80 |
| Eleven Realtime v1 Max | 77.3 | 0.57 s | $4.80 |
xAI's self-reported component scores for 3.0 include Big Bench Audio at 98.5%, Task Success at 94.8% and τ-Voice at 71.6%. On τ-Voice, GPT-Live-1 (Astra) is higher at 74.5%. The index lead comes mostly from Arena preference and task success, not from every axis.
Latency also got slightly worse: 3.0 averages 0.81 s to first audio against 0.70 s for 2.0, and ElevenLabs is faster still.
On
server_vadyou pay for session duration, so hold music, idle IVR legs and slow backend lookups all land on the bill.
Here is my illustrative arithmetic for unit economics. One 5-minute call on server_vad costs 5 × $0.08 = $0.40, including any hold time or silence while the session is open. At 10,000 such calls a month that is 50,000 min × $0.08 = $4,000. The concurrency ceiling is 10 sessions × 60 min = 600 call-minutes per hour, or 120 five-minute calls per hour per team until you get the limit raised. Applying AA's normalized hourly costs to a 5-minute call gives GPT-Live-1 (Astra) ≈ $5.83 ÷ 12 = $0.49 and Gemini 3.8 Live ≈ $3.50 ÷ 12 = $0.29. These are rough comparisons only: AA derives them from a fixed 40-question Big Bench Audio run, not from real call duration, and token-billed competitors scale with how much each side talks.
None of this includes telephony carrier fees, which you pay on top regardless of vendor.
Technical meaning and trade-offs
Flat pricing is a hedge, not automatically a saving. Per-minute billing removes token-count variance, which helps when calls run long or reasoning-heavy. But billing on session duration means idle IVR legs, hold music and slow backend lookups all cost money. A token-metered competitor charges little for silence.
Alias upgrades are a reliability event. A model change that moves time to first audio and changes turn-taking behaviour is something you want to canary, not wake up to. My read: anyone in production should ask xAI for a pinned model ID.
The benchmark is reasonable but thin on tails. AA reports average time to first audio. It doesn't report p95, measurement points or network conditions. Some competitor τ-Voice entries are based on a single trial. "Under one second" on average says little about the 5% of turns that make callers hang up.
What I'd do
First, pin or canary. If you're on grok-voice-latest, put a fraction of traffic on a fixed configuration and compare containment and handle time week over week. Second, measure p95 time to first audio yourself, from end of caller speech to first audio byte, over real SIP from your carrier's region to us-east-1.
Third, test barge-in: interrupt mid-sentence, interrupt during a tool call, talk over a confirmation, and check that playback stops and the transcript reflects what was actually said. Fourth, test tool calling during speech. Use parallel calls (the docs say not to send response.create until all outputs are in), slow tools (2 to 5 s), and failing tools, and listen for dead air or hallucinated confirmations.
Fifth, model your bill on session time, not talk time. Pull a week of call-detail records and add up connected minutes; consider push-to-talk or closing sessions during holds. Sixth, ask for a concurrency increase before launch. Ten sessions is a pilot, not a contact centre.
What I couldn't verify
The language count is unresolved: docs say "20+", product page "25+", and no list was checked. I also couldn't establish when and why the price moved from $0.05 to $0.08/min. xAI's customer figures (Starlink's 20,000 calls a day and ~70% resolved without a human) are self-reported. xAI's EVA-Bench claim (74% accuracy) has no competitor numbers on the page. And I found no p95 latency figure from any source.
I'll revisit when Artificial Analysis publishes tail latency or xAI exposes pinned voice model IDs.
Sources
Primary: xAI, Grok Voice 3.0; xAI, Grok Voice (API product page); xAI docs, Speech to Speech model page (pricing, limits); xAI docs, Voice agent capability guide; xAI, Grok Voice Agent API launch (Dec 2025). Benchmark: Artificial Analysis, Speech to Speech leaderboard.