xAI vs OpenAI Voice: Grok Voice 3.0 vs GPT-Live-1.1 Compared

On October 14, 2026 xAI released Grok Voice 3.0, a speech-to-speech model for its Voice Agent API, at the same $0.08 per minute as the Grok Voice 2.0 it replaces. On October 15 OpenAI's release notes announced GPT-Live-1.1 and GPT-Live-1.1 Quip, a smaller version, rolling out inside ChatGPT Voice with more natural accents and new voices in six languages. The two are not the same kind of product. Grok Voice 3.0 is something a business builds a phone agent on. GPT-Live-1.1 is something a ChatGPT subscriber talks to, and the model OpenAI sells to developers is still GPT-Live-1.

That distinction is the reason this post exists. Coverage of "xAI vs OpenAI voice" tends to put the two upgrades in one sentence and the two prices in one table. The prices are in different units with different inclusions, xAI's benchmark table compares Grok Voice 3.0 against GPT-Live-1 rather than 1.1, and OpenAI has given no date for 1.1 on its API. Each of those is a trap for a buyer, and each gets its own section below.

Sources are the vendors' own pages, read on October 16: xAI's announcement and pricing page, OpenAI's ChatGPT release notes, Voice help article and API pricing page, and Google's Gemini API pricing page for the third reference point. One naming note, once: xAI's announcement and pricing pages now carry "SpaceXSI" branding in places; this post uses xAI, the name in their x.ai and docs.x.ai addresses. Our earlier posts cover Grok Voice 2.0 and GPT-Live-1 on the API.

xAI's announcement of Grok Voice 3.0 describes a new speech-to-speech model that xAI calls "the world's most intelligent speech-to-speech model", a vendor claim we print as one. It replaces Grok Voice 2.0 in the Voice Agent API, and xAI states that agents pointed at the grok-voice-latest alias were switched over automatically with no changes needed. The pricing page, read October 16, lists Grok Voice 3.0 at $0.08 per minute, or $4.80 an hour, and marks 2.0 as deprecated.

OpenAI's ChatGPT release notes for October 15 say the company is rolling out GPT-Live-1.1 and a smaller GPT-Live-1.1 Quip in ChatGPT Voice, with more natural accents across languages, a wider range of the short phrases the model says while it searches, and new voices for French, German, Hindi, Japanese, Korean and Mexican Spanish plus a new English voice named Bubbie. Voices vary by region and account. Nothing in the notes mentions the API, and OpenAI's API pricing page on October 16 lists gpt-live-1 only.

The surface decides who the buyer is. Grok Voice 3.0 is reached through xAI's Voice Agent API, so its customer is a developer or a business wiring speech into a phone line, an app or a kiosk. xAI's announcement puts its own deployments forward as proof: Tesla vehicles handling what xAI describes as millions of conversations a day, and Starlink call centres taking about 20,000 calls a day, with roughly 70% resolved with no human and one in four sales calls ending in a purchase. All of those are xAI's figures about its deployments at Tesla and SpaceX, and we have no way to check them.

GPT-Live-1.1 is reached by opening ChatGPT and tapping the voice button. OpenAI's Voice help article describes three modes: Live, which is GPT-Live-1.1 or Quip depending on plan; Advanced, which OpenAI calls the previous real-time experience and which is still needed for video or screen sharing on mobile; and Standard, which is turn by turn. Spoken approval of plugin actions is not supported, so an action that needs approval is approved on screen. The customer here is a person or a workspace administrator choosing a plan, and the useful question is hours, not price per minute.

The failure case is a business that reads "OpenAI upgraded voice" and expects its gpt-live-1 phone agent to improve. It will not, until OpenAI ships 1.1 on the API, and on October 16 OpenAI had given no date for that. Secondary coverage saying the API version is "coming soon" is not an OpenAI statement.

Three vendors, three units. xAI prices Grok Voice 3.0 at $0.08 per minute of voice session; its pricing page also lists a line of "$0.004 / text input" whose unit we could not determine from the page, so we do not interpret it. OpenAI prices gpt-live-1 voice sessions at $0.05 per minute, billed per second, and states on the same page that "Backend model and tool usage is charged separately." Google prices Gemini 3.8 Live by token, at $3.00 per million audio input tokens and $12.00 per million audio output tokens including thinking, and gives its own per-minute equivalents of $0.005 in and $0.018 out.

Why the inclusions matter more than the headline: a gpt-live-1 session that calls a reasoning model and three tools behind the voice layer costs $0.05 a minute plus whatever those calls cost, which can be more than the voice line itself. xAI's $0.08 is presented as the rate for the voice agent; whether it covers the model's reasoning on long calls is not spelled out on the pricing page in a way we can quote. The example is a 10-minute support call: $0.80 on Grok Voice 3.0 by the published rate, $0.50 on gpt-live-1 before backend usage, and about $0.12 of audio on Gemini 3.8 Live if the caller and agent speak equal halves, by Google's per-minute figures. The failure case is calling any of those "cheapest" without the inclusions attached. Our guide to voice pricing units explains why minutes, tokens and characters do not convert.

For a ChatGPT subscriber, GPT-Live-1.1 is rationed in hours per rolling 24-hour window, and the help article updated around October 15 sets them out by plan. The table below is OpenAI's schedule as published; when the Live hours run out, ChatGPT notifies you, Business plans move to 1.25 credits a minute, and usage-based Enterprise pays per minute throughout. The article does not say what consumer plans fall back to.

Why it matters: a 20-person team that adopts ChatGPT Voice for daily stand-ups and dictation will hit 3 hours per person on Business Standard inside a working day of heavy use, after which every minute costs 1.25 credits. The example that surprises people is the $100 Pro tier: 15 hours, not unlimited; only the $200 tier removes the cap, and even that carries abuse guardrails. The failure case is a workspace administrator budgeting on seat price alone when the real variable is Live minutes past the included hours.

xAI's announcement carries a benchmark table citing third-party evaluations, and it compares Grok Voice 3.0 with GPT-Live-1 in its Astra configuration at medium effort, with Gemini 3.8 Live with extended thinking at high, and with its own Grok Voice 2.0. It does not compare with GPT-Live-1.1, which OpenAI released the day after xAI's announcement, and no shared benchmark covered that model on October 16. Read every row below as xAI's selection of third-party numbers against OpenAI's previous model.

Two readings. First, by xAI's selection Grok Voice 3.0 leads every row, and the gap over GPT-Live-1 ranges from about two points on Full Duplex Bench to eight points on Big Bench Audio and 22 Elo points on the arena. Second, the improvement over xAI's own 2.0 is uneven: the index rises from 81.3 to 86.9 and the arena from 1011 to 1070, while task success barely moves from 94.6% to 94.8%. For a buyer already on 2.0 that says the upgrade is about comprehension and conversational quality, not about completing more tasks. We have not run any of these; our reference on voice-agent latency measures is the method for checking the claims on your own calls.

Beyond the table, xAI makes three claims that a buyer will hear repeated. On τ-Voice, which xAI says measures support calls with varied accents, noise and interruptions, xAI reports Grok Voice 3.0 at 71.6% and describes it as comparable with GPT-Live-1 in its Astra configuration but at 40% lower latency. On EVA-Bench, an evaluation from ServiceNow, xAI reports Grok Voice 3.0 at 74% accuracy and best jointly on accuracy and customer experience, and carries a ServiceNow quote saying it "sets a new Pareto frontier on EVA-Bench". And xAI describes response latency as under one second. EVA-Bench is ServiceNow's published open-source benchmark, but xAI gives no method for the 40% latency ratio or the sub-second figure, and the ServiceNow line is a partner quote on the vendor's page.

Why it matters: latency is the number a caller feels, and it is the one nobody in this comparison publishes as a distribution. "40% lower" than an unstated figure is a ratio without a base. The example is a support line where the agent must look up an order before answering: tool latency, not model latency, usually dominates, and no vendor table measures your tools. The failure case is choosing on a latency claim and discovering that the database call is the slow part. Our Gemini 3.8 Live post covers the same trade-off from Google's side, where extended thinking buys accuracy with wait time.

The routing below follows the surfaces, the published prices and xAI's cited benchmarks. It is where to start a two-week pilot, not where to end one.

If your team is choosing a voice model for a customer line and wants the three APIs tested on your own call recordings, tools and latency budget rather than on a vendor's table, that is work our AI transformation team does, including the per-call cost model with every inclusion written down.