Grok Voice 3.0 Is First on the Speech Index at the Same $0.08 a Minute. It Did Not Win a Single Column.

In May we wrote about Grok Voice Think Fast 1.0, the $0.05-a-minute speech model that reset the voice-agent benchmark. The "Think Fast" name is gone from xAI's docs, the company has changed its own name twice since, and the model is on its third major version. Today, October 14, 2026, SpaceXSI announced Grok Voice 3.0: "the world's most intelligent speech-to-speech model. It sets a new record for resolving customer support calls, has class-leading latency at under one second, and costs less than other frontier models."

The independent half of that holds. Artificial Analysis's Speech to Speech Index, which xAI cites as of October 13, puts Grok Voice 3.0 first at 86.9, ahead of OpenAI's GPT-Live-1 running Astra at 83.2 and Google's Gemini 3.8 Live Extended Thinking at 82.6. The price is $0.08 per minute, exactly what Voice 2.0 cost. And if your agent was pinned to grok-voice-latest, it is already running 3.0: "Agents using grok-voice-latest have already been switched over automatically, with no changes needed."

The rest needs the asterisks. First place on the index comes from being near the top of all four equal-weighted components, not from winning them; on the support-call benchmark that matters most to the Starlink use case xAI leads with, GPT-Live-1 is ahead; and 3.0 is slower to first audio than 2.0. We read the announcement, the pricing page, the release notes, the voice-agent and SIP docs, AA's board, and OpenAI's and Google's price pages. Here is what held up.

Same Price, Same API, New Model Under Your Alias

The alias switch is the fine print. The docs: "grok-voice-latest points to grok-voice-3.0, so sessions that use the alias already run on Grok Voice 3.0." There was no opt-in and no overlap window. If you pinned grok-voice-2.0 you still have it, deprecated, for an unstated period. The API itself did not change: same endpoint, same session parameters, same events.

One price covers the session. $0.08 a minute, audio in and out, with text input at $0.004 per message. That is simpler than the rivals' sheets, and it matters for the comparison below, because "cheaper than OpenAI" depends entirely on which line of OpenAI's sheet you read.

No card. There is no model card or system card for Grok Voice 3.0, so there are no safety figures to print, and the capability numbers below are AA's or xAI's.

The Benchmarks, Sorted by Who Reported Them

House rules: independent numbers first, stated as fact with the source and a date, because AA's speech board drifts; xAI's July post cited Voice 2.0 at 82.9 and Voice 1.0 at 75.7 on the same index, and the board today prints 81.3 and 73.7 for the same models.

Artificial Analysis Speech to Speech Index, October 13. The index is an equal-weighted average of four components: speech reasoning (Big Bench Audio), human preference (Speech Agent Arena Elo), task completion (Task Success Rate), and agentic performance (tau-Voice). Grok Voice 3.0 scores 86.9.

Where the 86.9 comes from, column by column: Task Success Rate 94.8%, first. tau-Voice 71.6%, second to GPT-Live-1 (Astra) at 74.5%. Speech Agent Arena 1070 Elo, third, behind Gemini 3.1 Flash Live Minimal at 1100 and Gemini 3.8 Live at 1088. Big Bench Audio 98.5%, fourth, behind StepAudio 3 Realtime at 99.7%, Qwen Audio 3.0 Realtime Plus at 99.2%, and Qwen3.5 Omni Plus Realtime at 98.7%. Full Duplex Bench, the turn-taking test, is 97.0% and is not an index component. One first, one second, one third, one fourth, and the average is the best average on the board. That is a real result; it is not the result "the world's most intelligent voice model" implies.

The gain over 2.0 is almost entirely one column. Voice 2.0 to Voice 3.0 moves tau-Voice from 56.5% to 71.6%, a 15-point jump on the support-call benchmark, while the other three components move a point or two. And 3.0's time to first audio, 0.81 seconds, is slower than 2.0's 0.70. Both are under one second; "class-leading latency" is a comparison with OpenAI and Google (1.34 and 1.35 seconds), not with xAI's own previous model.

Now xAI's numbers. The announcement's comparison table is "per Artificial Analysis" and its cells match AA's for the Grok models; two rival cells do not (xAI prints GPT-Live-1's Arena Elo as 1048 and Gemini 3.8 Live's as 990; AA's board shows 1056 and 993). xAI's tau-Voice sentence is "Grok Voice 3.0 scores 71.6%, comparable with GPT-Live-1 (Astra), but at 40% lower latency." Comparable is a generous word for second. xAI's cost-per-task chart, from AA's Big Bench Audio runs, shows $0.025 for Grok Voice 3.0 against $0.030 for GPT-Live-1 (Astra, medium) and $0.055 for GPT-Realtime-2.1 (high).

Three honest observations:

• "Costs less than other frontier models" is true on measured cost and false on list price. OpenAI's gpt-live-1 lists at $0.05 per minute, billed per second, which is less than $0.08; the backend model and tool usage are billed separately, which is why AA's measured all-in cost comes out at $5.83 an hour against Grok's $4.80. Google's Gemini 3.8 Live lists at $0.005 a minute in and $0.018 out and measures at $3.50 an hour, cheaper than Grok on both bases. AI Weekly's "17-55% below comparable OpenAI offerings" is wrong on the list price and shows no arithmetic.
• EVA-Bench is a one-model leaderboard. xAI says Grok Voice 3.0 "achieves the highest accuracy score at 74%" on ServiceNow's enterprise voice-agent benchmark, with a ServiceNow director quoted calling it "a new Pareto frontier." ServiceNow publishes the benchmark and no leaderboard, so there is no rival score to print. Dash.
• The deployment claims predate the model. "Starlink call centers use Grok Voice to answer 20,000 customer calls every day," with "roughly 70% resolved with no human in the loop" and "one in four sales calls" converting, says Grok Voice, not Grok Voice 3.0, and the 70% figure appeared in xAI's materials before today. The Tesla line, "millions of conversations a day," is likewise unversioned. Vendor claims, not independently measurable.

Where Grok Voice Came From, and Where 4.8 Is

The lineage since our May post: Grok Voice Think Fast 1.0 in April at $0.05 a minute; Grok Voice 2.0 on July 29 at $0.08, when the "Think Fast" branding disappeared; Grok Voice 3.0 today at the same $0.08. No think-fast model ID remains in the docs. AA's board still carries Grok Voice 1.0 at 73.7 and a "Grok Voice Fast 1.0" with no score.

Two smaller xAI items from the same fortnight. Grok Imagine Video 1.5 Lite is the budget tier of the video model: text-to-video and image-to-video with native audio, 1 to 15 second clips, priced by resolution at $0.02 per second for 480p, $0.03 for 720p, and $0.14 for 1080p, plus a cent each for text and image input; the full 1.5 model is $0.08, $0.14, and $0.25. It appeared on fal on October 1 and Runware on October 2; a newsletter's "announced October 8" is contradicted by both. On AA's text-to-video board the Lite model ranks 17th at 999 Elo and the full 1.5 tenth at 1060.

Grok 4.8 is still a tweet. Musk on September 14: "Grok 4.8, which is a 2.5T model trained with our new C++ software stack, will finish training this week and start RL." As of today it is absent from the pricing page, the models list, the release notes, and Cursor's model docs; no date has come from xAI or Musk, and the "October 2026" on Wikipedia is cited to a betting-odds tracker. On October 7 Musk said most Grok Bot requests "will be handled by a lightning-fast version of Grok 4.8 when that comes out." Meanwhile Grok 4.7, which our September post noted was not in the consumer app, appears to have reached grok.com and the Grok app around October 1 to 3 according to two third-party reports; xAI has said nothing either way.

The company is now SpaceXSI, "SpaceX Super Intelligence," its third name this year after xAI and SpaceXAI, following an executive order that told federal agencies to say "Super Intelligence" instead of "AI." The footer of x.ai and AA's provider label still say SpaceXAI.

Where You Can Run It

The Voice Agent API and the Voice Agent Builder in the xAI console, in three regions. First-party at Tesla and Starlink. Not on Bedrock, Vertex, Foundry, or OCI. Whether the consumer Grok app's voice mode now runs 3.0 is not stated anywhere; dash.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use Grok Voice 3.0 if you run voice agents that complete tasks, where it is first on AA's Task Success Rate, want sub-second first audio with one flat price, need SIP telephony and MCP tools in the same session, or already run Voice 2.0, in which case you are using it whether you chose to or not. For support calls, it is the second-best agent on the board at 82% of GPT-Live-1's measured cost.

Skip it if your agent's job is the support call itself and the last three points of task completion matter; GPT-Live-1 running Astra leads tau-Voice. Skip it if measured cost is the whole decision, because Gemini 3.8 Live Extended Thinking scores within four index points at $3.50 an hour. And pin a model ID: the alias will move again.

The frontier is elsewhere: on text models Grok 4.7 sits at 46 on AA's Intelligence Index against Claude Opus 5.5's 58, and Grok 4.8 has a parameter count and no date. Voice is the one board where SpaceXSI is first today. Our model comparison has the current standings side by side.

Update, October 16, 2026

• October 15: OpenAI moved ChatGPT Voice to GPT-Live-1.1, which is ChatGPT-only with no API model, price, or benchmark; the API's gpt-live-1 is unchanged at $0.05 a minute, so the AA comparison above still stands.
• October 16: AA's board is unchanged at the top. No release-note entry, price change, outage, or correction has followed the launch. Google's models page lists Gemini 3.8 Live as a stable release, not a preview, whatever some coverage says. The consumer Grok app's voice version is still unstated, and Voice 2.0 still has no shutdown date.

Sources

• SpaceXSI: Introducing Grok Voice 3.0
• docs.x.ai: pricing
• docs.x.ai: release notes
• docs.x.ai: Voice Agent API
• docs.x.ai: Speech to Speech API and upgrading from Voice 2.0
• docs.x.ai: SIP telephony
• docs.x.ai: rate limits
• Artificial Analysis: Speech to Speech Index
• OpenAI model docs: gpt-live-1
• docs.x.ai: grok-imagine-video-1.5-lite

*Not sure whether a voice agent belongs in your stack, or which lab's version fits the calls you actually take? Take the free 2-minute quiz and get matched. Then read about the rival that returns a talking face as well as a voice, with a per-minute bill of its own: Gemini 3.8 Live with Live Avatar.*