On October 14, 2026, xAI shipped a voice model and buried the most interesting thing about it near the bottom of its own announcement. Grok Voice 3.0 debuted at #1 on Artificial Analysis's Speech-to-Speech Index with a score of 86.9, a 94.8% task success rate, and an API price of $0.08 per minute. It is also already in production: it answers the phone at Starlink and talks to drivers inside Tesla cars, resolving roughly 70% of inbound support calls with no human on the line and closing a sale on something like one in four sales calls. The Hacker News thread about all of this reached six points and two comments, one of which read, in full, "hackernews is sleeping on this model." That is funny, because this is the least speculative AI story of the month — not a demo, not a benchmark stunt, but the thing that picks up when you call about your dish. The problem is that the numbers describing it do not match from one source to the next.
Nobody Agrees on How Many Calls This Thing Takes
xAI's launch page puts Starlink at 20,000 customer calls a day. A write-up at Dapta covering the same deployment puts it at "over 15,000" inbound calls daily. Same system, same company, same job, two numbers 5,000 calls apart — and neither source labels the measurement window or says whether chat traffic is folded in.
| Metric | xAI's launch page | Dapta's write-up |
|---|---|---|
| Daily Starlink calls | 20,000 | 15,000+ inbound |
| Resolved with no human | ~70% | ~70% |
| Sales calls that convert | 25%, before the caller disconnects | ~20% |
| Orders fulfilled | not stated | 3,000+ per week, voice and chat combined |
| Independent audit | not mentioned | explicitly none published |
| When callers first heard an AI | not stated | mid-April 2026 |
The Gap Is a Third of the Smaller Number
There is a benign explanation available. Dapta reports that Starlink callers began hearing the system identify itself as AI in mid-April 2026, which means the deployment had been scaling for about six months by the time Voice 3.0 landed. A figure captured in spring and a figure captured in October would naturally differ, and 15,000 to 20,000 over half a year is an unremarkable growth curve. The conversion rates split the same way — 20% versus 25% — which is consistent with two snapshots rather than two contradictions. But nobody dates either figure, so a reader cannot tell growth from disagreement. When a company is the sole source for a number, the date on it is doing as much work as the number.
"70% Resolved" Is a Self-Reported Metric
Dapta is unusually direct about this, noting that the figures come from the company itself and that no independent audit of resolution rates or caller satisfaction has been published. That caveat should travel with the number everywhere it goes. "Resolved" is also undefined in every source: a caller who gives up and hangs up has ended the call without a human, and whether that lands in the 70% or outside it changes what the statistic means entirely. Note the phrasing xAI chose for the sales figure too — conversion happens "before the caller disconnects." That is not marketing fluff; it is a disclosure about where the measurement boundary sits.
The Benchmark Sheet, Read Slowly
The headline is real: 86.9 on the Speech-to-Speech Index puts Grok Voice 3.0 ahead of GPT-Live-1 (Astra, Medium) at 83.2 and Gemini 3.8 Live (Extended Thinking, High) at 82.6. But the composition of that lead is more interesting than the lead itself.
| Model | Speech-to-Speech Index | Speech reasoning (Big Bench Audio) | Human preference (Elo) | Task success | Full Duplex Bench |
|---|---|---|---|---|---|
| Grok Voice 3.0 | 86.9 | 98.5% | 1070 | 94.8% | 97.0% |
| GPT-Live-1 (Astra, Medium) | 83.2 | 90.1% | 1048 | 87.4% | 94.9% |
| Gemini 3.8 Live (Ext. Thinking, High) | 82.6 | 97.7% | 990 | 89.1% | 91.9% |
| Grok Voice 2.0 | 81.3 | 97.2% | 94.6% | 94.6% | 95.1% |
Against Its Own Predecessor, Task Success Barely Moved
Grok Voice 2.0 already scored 94.6% task success. Version 3.0 scores 94.8% — two-tenths of a point. The index gain from 81.3 to 86.9 comes from elsewhere: speech reasoning 97.2% to 98.5%, Full Duplex Bench 95.1% to 97.0%, and human-preference Elo from 1011 to 1070, which is the biggest single jump on the sheet. Read plainly, 3.0 is a model people like talking to more, not a model that completes more of the simple stuff. If your deployment's bottleneck is completion rate on routine requests, the upgrade may not move your dashboard at all.
The Agentic Number Jumped — and Still Loses
The one benchmark where 3.0 is a different animal is τ-Voice, which measures agent success on multi-step voice tasks: 56.5% for Voice 2.0, 71.6% for Voice 3.0. Fifteen points in one generation is a serious result, and it is the number that explains why this model can be trusted with a support queue at all. It is also the number where the #1-ranked model is not first. GPT-Live-1 (Astra, Medium) scores 74.6% — call it 74.5% as reported — roughly three points ahead. Gemini 3.8 Live sits at 68.6%. So "ranked #1 on the speech index" and "best at getting an agentic voice task done" are two different claims, and only one of them is xAI's to make.
The Caveat the Press Releases Skipped: It May Have Gotten Slower
Eastern Herald's reading of the same Artificial Analysis data reports time to first audio at 0.81 seconds for Voice 3.0 against 0.70 seconds for Voice 2.0 — about a tenth of a second given back to the clock in exchange for the higher index score. xAI's own chart plots the two models differently, placing 3.0 at roughly 0.4 seconds and 2.0 at roughly 0.7. Those two accounts cannot both be right, and neither side publishes the test conditions that would settle it, so treat the version-over-version latency as unresolved rather than decided.
What survives either reading is the competitive picture. Against rivals, Grok Voice 3.0 is still the fast one — GPT-Live-1 comes in around 1.3 seconds to first audio and Gemini 3.8 Live around 1.9 seconds — and xAI's claims of sub-second response and 40% lower latency than GPT-Live-1 hold up on either number. The honest one-line summary is awkward for a launch post, which is presumably why nobody wrote it: faster than everything else on the board, and possibly slower than the model it replaced.
Why a Tenth of a Second Is a Real Decision
If the Eastern Herald figures are right, xAI traded 0.11 seconds of opening silence for 5.6 points of index and 15 points of τ-Voice agent success. For a refund request or an address change, that is an obviously good trade — the caller would rather wait a beat and get the right answer. For anything where barge-in and interruption timing carry the experience, it is a trade you should measure on your own recordings instead of taking on faith. The models are close enough now that the question is no longer which one is smarter but which one's failure mode your callers tolerate.
The Price Did Not Move, Which Is the Actual Story
Grok Voice 3.0 lists at $0.08 per minute — $4.80 per hour — identical to Voice 2.0, with text input billed separately at $0.004 per input. On cost per task on Big Bench Audio, xAI reports $0.025 against $0.030 for GPT-Live-1 (Astra, Medium) and $0.055 for GPT-Realtime-2.1 (High).
| Model | API price per minute | Cost per task (Big Bench Audio) | Time to first audio | τ-Voice agent success |
|---|---|---|---|---|
| Grok Voice 3.0 | $0.08 | $0.025 | ~0.4s (xAI) / 0.81s (Eastern Herald) | 71.6% |
| Grok Voice 2.0 | $0.08 | not reported | ~0.7s | 56.5% |
| GPT-Live-1 (Astra, Medium) | not disclosed in sources | $0.030 | ~1.3s | 74.5% |
| GPT-Realtime-2.1 (High) | not disclosed in sources | $0.055 | not reported | not reported |
| Gemini 3.8 Live (Ext. Thinking, High) | not disclosed in sources | not reported | ~1.9s | 68.6% |
This Is Where the Call-Volume Dispute Starts Costing Money
Attach the per-minute rate to the deployment and the vagueness gets expensive. Assume a five-minute support call — an assumption, not a sourced figure — and one call costs $0.40 of inference. At xAI's 20,000 calls a day that is $8,000 daily, roughly $2.9 million a year. At Dapta's 15,000 it is $6,000 a day, about $2.2 million. The $700,000 spread between those two is entirely a function of which press number you believe, which is a decent illustration of why anyone modelling a voice deployment should ignore published volumes and price their own traffic. For a smaller shop the arithmetic is friendlier and just as blunt: 500 five-minute calls a day runs $200, or about $6,000 a month, before you count a single escalation to a human.
The Alias Swap That Broke Pinning
The operational story landed on the same day as the benchmarks. On October 14, the mutable grok-voice-latest alias stopped pointing at grok-voice-2.0 and started pointing at grok-voice-3.0, with Voice 2.0 marked deprecated. Anyone running a voice agent on the alias got a new model in production with no deploy, no config change, and no announcement in their own change log.
The DEV Community write-up on this is worth reading in full if the alias is in your codebase, because it enumerates what actually shifts underneath a live phone system: scripts may interrupt callers sooner or hold a beat longer, escalation rates to human agents can move, tool and web-search calls fire at different moments and therefore cost different amounts, and instruction-following behavior changes in ways a prompt tuned against 2.0 will not anticipate. None of those are bugs. They are the normal consequence of swapping the model out from under a tuned system, which is precisely why you pin.
The Five-Minute Checklist
- Grep the codebase for grok-voice-latest and every config, environment file, and infrastructure template that might carry it.
- Pin explicitly to grok-voice-3.0 or grok-voice-2.0 — choose a version rather than inheriting one.
- Replay recorded call scenarios against the pinned version before promoting it, with interruption timing specifically in scope given the unresolved latency question.
- Log the exact model ID on every session so a future regression is attributable instead of mysterious.
- Add a CI check that fails any -latest alias in a production config.
Pinning is also the only way to answer the 0.4-versus-0.81-second question for yourself. Run both versions against the same recordings and measure your own time to first audio; it is a better number than either published one because it is measured on your traffic.
What Is Still Unknown
A fair amount, and the gaps are specific. There is no independent audit of the 70% resolution rate or of caller satisfaction at Starlink. There is no published definition of "resolved." The daily call volume is reported two ways and dated neither. The version-over-version latency is reported two ways and tested publicly by nobody. The corporate shape has moved too — Dapta reports that SpaceX absorbed xAI earlier in 2026, and the launch post on X came from an account styled SpaceX Super Intelligence, calling Voice 3.0 the world's most intelligent speech-to-speech model that "excels at resolving customer calls." Elon Musk's framing on X, that it is the best conversational voice model in the world, is a claim about a composite index on which the model is first overall and third on agentic task success.
The useful takeaway is not about xAI at all. A frontier voice model is now cheap enough at $0.08 a minute, fast enough at under a second, and good enough at 94.8% task success that a company with a serious support load has already shipped it to tens of thousands of real callers a day and stopped treating it as an experiment. That threshold has been crossed. What has not arrived is any independent accounting of what happens on those calls — and until a third party audits a deployment at this scale, every resolution rate you read, including the one in this post's headline, is a company describing itself. Pin your model, log your sessions, and measure your own callers.
-EditorZ
Photo by Siwawut Phoophinyo on Unsplash
Post a Comment