Grok Voice 3.0 leads Artificial Analysis’s speech index at eight cents a minute

A support call that survives interruption, accent shift, and a tool call is the product people actually buy. On 14 October 2026, SpaceXSI put Grok Voice 3.0 on that job at eight cents a minute and pointed to an independent scoreboard.

Artificial Analysis’s Speech-to-Speech Index, dated 13 October in SpaceXSI’s table, puts the new model at 86.9. That sits ahead of OpenAI’s GPT-Live-1 in its Astra medium configuration and Google’s Gemini 3.8 Live. The list price did not move from Grok Voice 2.0. If your agent already called grok-voice-latest, you are already on it.

Grok Voice 3.0 ships on the same realtime WebSocket SpaceXSI already sold

Grok Voice 3.0 is a native speech-to-speech model for the Voice Agent API. You open a WebSocket to wss://api.x.ai/v1/realtime, stream audio both ways, and configure voice, instructions, turn detection, and tools in a session update. SpaceXSI’s docs still describe the surface as compatible with the OpenAI Realtime event shape: change the base URL, the key, and the model string.

That continuity matters more than a new SDK. The company says the endpoint, session parameters, and events are unchanged from Grok Voice 2.0. Built-in voices and custom voice IDs carry over. Only the weights behind the socket moved.

Artificial Analysis puts Grok Voice 3.0 at 86.9 on the Speech-to-Speech Index

The Speech-to-Speech Index equally weights speech reasoning on Big Bench Audio, agentic performance on τ-Voice, human preference in live arena calls, and task success rate. Grok Voice 3.0 sits at 86.9. Grok Voice 2.0 was 81.3. GPT-Live-1 (Astra, Medium) is 83.2. Gemini 3.8 Live (Extended Thinking, High) is 82.6.

The component rows SpaceXSI published from that run are blunt. Speech reasoning hits 98.5%. Preference Elo is 1070. Task success is 94.8%. Full Duplex Bench, which scores pause handling, turn-taking, interruptions, and backchannels, lands at 97.0%.

τ-Voice is the slice that leaves the lead unfinished. The benchmark walks support calls through accents, noise, and interruptions. Grok Voice 3.0 scores 71.6%. Reporting around the launch puts Astra at 74.5% on the same measure. SpaceXSI answers with latency: about 40% lower time to first audio on that test. Treat the latency percentage as the company’s reading of the Artificial Analysis run until AA’s public pages show the same pairing.

ServiceNow’s open-source EVA-Bench adds a thinner signal. SpaceXSI says Grok Voice 3.0 scored 74% accuracy across more than 200 enterprise scenarios and led when accuracy and customer experience were scored together. That is a vendor-cited partner bench beside the Artificial Analysis board, useful colour until you re-run the scenarios yourself.

Eight cents a minute, and grok-voice-latest already cut over

Pricing is the quiet part of the ship. Grok Voice 3.0 is $0.08 per minute, or $4.80 per hour, the same as Grok Voice 2.0. SiliconANGLE’s write-up of Artificial Analysis’s Big Bench Audio cost chart puts average cost per task near 2.5 cents on Grok Voice 3.0, against about 3 cents on Astra and 5.5 cents on GPT-Realtime-2.1. Those figures come from secondary reporting of AA’s chart.

Availability is already partially decided for you. grok-voice-latest now resolves to grok-voice-3.0. SpaceXSI says agents on the alias were switched automatically. grok-voice-2.0 is deprecated. The docs still advise pinning a versioned model name in production if you want change control. The API looks production-shaped. The internal Tesla and Starlink loads remain vendor testimony.

The AA composite and τ-Voice still split the Monday call

Teams already on SpaceXSI’s realtime socket have the cheapest Monday test. Point a staging agent at grok-voice-3.0, keep a pinned 2.0 or an Astra control if you still have access, and listen for interruption handling on the calls you actually take. The migration cost is a model string.

If the Artificial Analysis index or Full Duplex Bench is the figure you have to defend, Grok Voice 3.0 is the row to beat this week at a flat eight cents. Contracts that hinge on τ-Voice-style support calls with heavy accents and background noise still face Astra’s published lead. Pin grok-voice-3.0 rather than grok-voice-latest when a surprise cutover would break a change window.

Tesla and Starlink volume claims remain SpaceXSI’s alone

“The world’s most intelligent voice model” is launch language. The Artificial Analysis index is the harder claim: a composite score on a defined harness.

The production story is louder and softer at once. SpaceXSI says Grok Voice handles millions of Tesla conversations a day, and that Starlink call centres take about 20,000 Grok Voice calls daily across languages, with roughly 70% resolved without a human and one in four sales calls ending in a purchase before hang-up. None of those figures arrives with an external audit in the launch packet. Until they do, they are company operating claims, useful as colour and dangerous as capacity planning inputs.

Two measurement gaps sit beside the marketing. τ-Voice still prefers Astra on the published numbers. The “about 40% lower latency” line is SpaceXSI’s framing of the same Artificial Analysis work. Average cost-per-task charts also diverge from your minute bill once tool calls, retries, and hold music enter the session.

Interrupt one of your own calls before you trust anyone’s resolution rate

Speech agents are now sold like infrastructure minutes with a public leaderboard attached. Grok Voice 3.0 owns the Artificial Analysis Speech-to-Speech Index this week at the same eight cents Grok Voice 2.0 already charged. Put one interrupted support call on Grok, Astra, and Gemini against your own accents. That listening test still outweighs SpaceXSI’s Starlink percentage.