Grok Voice 3.0 Tops OpenAI and Google on Voice Benchmark

Grok Voice 3.0 Tops OpenAI and Google on Voice Benchmark

SpaceXSI, the AI company owned by Elon Musk's SpaceX, has released Grok Voice 3.0. It is a new speech-to-speech model that powers the company's Voice Agent API, and it now sits at the top of an independent ranking ahead of models from OpenAI and Google.

Readers who have lost track of the name: SpaceXSI is the former xAI. SpaceX acquired xAI in February. The AI unit became SpaceXAI in July and took its current name earlier this month.

The target market is narrow and practical. SpaceXSI wants businesses that run customer support and sales lines through voice agents. These are systems that listen, reason and talk back in real time, with no human operator on the line.

How the benchmarks stack up

The main claim comes from Artificial Analysis, an independent firm that tests and ranks AI models. Its Speech-to-Speech Index combines three inputs: reasoning tests, task success rates, and votes from listeners who compare two models side by side in blind calls.

On that index, the scores were:

  1. Grok Voice 3.0: 86.9
  2. OpenAI's GPT-Live-1: 83.2, tested with OpenAI's GPT-6 Astra model doing the reasoning
  3. Google's Gemini 3.8 Live: 82.6

The new Grok model also scored 97% on Full Duplex Bench. That test measures how naturally a model handles pauses and interruptions, which is where many voice agents start to feel robotic.

Grok did not win everywhere. On Tau-Voice, a test built around support calls with heavy accents and background noise, OpenAI's model held a small lead at 74.5% against Grok's 71.6%. SpaceXSI says its model produced its first audio about 40% faster on that test. Speed matters on a phone line, but it is a separate question from getting the answer right.

Cost is another part of the pitch. In Artificial Analysis' Big Bench Audio run, a single task cost about 2.5 cents on Grok Voice 3.0. The same task cost 3 cents on GPT-Live-1 and 5.5 cents on OpenAI's GPT-Realtime-2.1.

An enterprise test from outside the lab

ServiceNow, the US enterprise workflow software company, maintains an open-source benchmark called EVA-Bench. It covers more than 200 enterprise scenarios. Grok Voice 3.0 led on accuracy there with 74%, and no model ranked higher when accuracy and customer experience were scored together.

Nitin Aggarwal, a senior director of product management at ServiceNow, said the model "pairs top accuracy with a natural, responsive conversation."

Already running at scale

SpaceXSI can point to deployments inside the wider Musk group of companies.

  • Tesla: Vehicles already handle millions of Grok Voice conversations a day. Drivers use it to check calendars and answer email. Larger tasks, such as booking dinner, go through Grok Bot, whose voice mode also runs on the new model.
  • Starlink: At SpaceX's satellite internet business, Grok Voice takes about 20,000 customer calls a day in several languages. According to SpaceXSI, roughly 70% are resolved without a human getting involved. One in four sales calls ends in a purchase before the caller hangs up.

These figures come from SpaceXSI, not from an outside audit. Both customers also belong to the same corporate family as the vendor.

Pricing and availability

Grok Voice 3.0 is available now in the Voice Agent API for eight cents a minute, or $4.80 an hour. That is unchanged from Grok Voice 2.0. Developers whose agents use the grok-voice-latest model alias have already been switched to the new version without changing anything on their end.

Our Take

Voice is turning into its own competitive front in AI, separate from chatbots and coding models. The numbers here show why. The gaps between the top three systems on the main index are a few points. On the noisy, accented call test, the ranking flips. For a business choosing a vendor, the lesson is that a single headline score says little about how a model will cope with its own callers.

The pricing and auto-upgrade details may matter more than the rankings. Better scores at the same per-minute rate, delivered silently through a model alias, lower the switching effort for existing customers. That auto-upgrade has a second side, though. A team that has tuned its call flows around one model now runs on another without having tested it. Teams with strict testing needs may want to pin a specific version instead of relying on "latest".

The Starlink figures point to the same tension. A 70% resolution rate without humans is a strong efficiency number. It also means the agent is the final word for most of those callers, including when it gets things wrong. As voice agents handle sales and support on their own, the question of who carries the blame for errors is getting more attention, and insurers are already preparing for liability claims tied to AI agents.

The market is moving in this direction from several sides. Automation Anywhere recently bought Boost.ai to expand into voice AI, which suggests enterprise vendors see voice as a core channel for agents, not an add-on.

It is worth watching whether OpenAI and Google answer with pricing changes or new real-time models. Another open point is whether third-party customers outside the Musk companies publish results comparable to Starlink's. Until they do, the strongest evidence for Grok Voice 3.0 comes from benchmarks and from deployments inside its own corporate group.