Introducing Grok Voice 3

Jessica Grok Voice 3.0 just shipped, and the thing that got my attention is they held the price. Same eight cents a minute as the last version.

Cathy Mm-hm.

Jessica That's a choice. They could have gone up, benchmarks are better, and they didn't. That feels like the real move to me — not the index score.

Cathy Okay, but let's talk about that index score first, because it's doing a lot of work in the announcement. Eighty-six point nine on Artificial Analysis's Speech-to-Speech Index. Sounds clean.

Jessica Right.

Cathy But the index is three different things mashed together — speech reasoning, τ-Voice, and preference scores from blind calls. It's not one measure. It's an aggregate, and xAI's announcement led with the aggregate number.

Jessica So you're saying the fifty-thousand-foot view looks great, but the component scores are what actually matter.

Cathy Exactly. τ-Voice is the one I want to see — how it handles real diversity, accents, background noise, interruptions in actual customer calls. Grok Voice 3.0 scores seventy-one point six percent. GPT-Live-1 scores the same. BUT Grok gets there forty percent faster — latency is zero point eight seconds versus one point three four.

Jessica Okay, that's genuinely useful.

Cathy It is, but here's the part I don't buy: they're claiming this resolves customer support calls. What does that mean? Does the agent complete a transaction? Does the caller say they're satisfied? Does it just mean the call didn't drop?

Jessica That's the thing — the benchmarks don't tell you. But Starlink is running this at scale, twenty thousand calls a day, and they're saying seventy percent resolve with no human in the loop. One in four sales calls closes before the caller hangs up.

Cathy Yeah.

Jessica That number is real. That's not a benchmark, that's a production system running millions of conversations, and it's actually working. The agent verifies customers, troubleshoots hardware, issues credits, ships replacement equipment — all hands-free.

Cathy That's the only number in the announcement I trust. Everything else is index theater.

Jessica But here's what I keep coming back to — xAI could have raised the price. They improved the model, the benchmarks are better, latency dropped. They held eight cents a minute. And they auto-upgraded existing users, no migration, no billing surprise.

Cathy That's interesting.

Jessica It's a product move, not a pricing move. You ship the better thing at the same cost, you make migration invisible, and you let the Starlink result speak for itself. That's how you get adoption without theater.

Cathy So from a competitive angle — this lands between GPT-Live-1, which is faster on some measures but costs more per task, and Claude Sonnet Voice, which is cheaper but doesn't have the end-to-end reasoning. Grok Voice 3.0 is trying to split the difference: latency, task completion, same price as before.

Jessica Right. And Tesla's using it for in-car stuff — Connectors for simple requests, Grok Bot for bigger tasks. That's a real deployment, not a demo.

Cathy I'd put seventy-thirty odds that within a month we see a public post from someone running both Grok Voice 3.0 and GPT-Live-1 side-by-side on their support queue, and the latency difference actually moves their call completion rate.

Jessica That's fair. The benchmark win is clean, but the real test is whether the latency drop changes what people automate.

Cathy Exactly.

Jessica This is one of those moments where the boring thing — same price, auto-upgrade, production proof — is the actual story.

Cathy Yeah. The Starlink number does more work than all the benchmarks combined.

Jessica Grok Voice 3.0 is live in the API right now, same pricing tier. If you're already using it, you're upgraded already. No friction.