Grok Voice 3.0 vs GPT-Live-1: xAI Won the Index and Lost the Test That Matters

Two flagships, 3.7 points apart on Artificial Analysis's Speech to Speech Index, and both of them beaten on the one component that decides whether a voice agent is worth deploying. Grok Voice 3.0 reads 86.9 on the index that SpaceXAI announced on 14 October 2026, against 83.2 for GPT-Live-1 in its Astra medium configuration, the version O​penAI's own docs point a WebRTC integration at when a session delegates reasoning to a backend. On τ-Voice — the customer-service benchmark that scores whether an agent actually completes a multi-step job — Grok Voice 3.0 lands at 71.6% and GPT-Live-1 Astra medium lands at 74.5%. The model with the best composite in the category does not hold the component that correlates with getting work done, and O​penAI did not publish a headline about it either.

The second thing worth noticing is the price direction. GPT-Live-1 bills $0.05 per minute of voice, charged by the second, which is $3.00 an hour. Grok Voice 3.0 bills $0.08 per minute, which is $4.80 an hour. The newer model costs 60% more per minute of audio, and it is the model whose vendor spent the last quarter of its predecessor's life absorbing a rate increase rather than announcing a cut.

The architecture split nobody puts on the comparison chart

Neither of these is a speech model that thinks. Both are voice layers that delegate, and the two delegation designs are not the same animal — which is why the integration cost differs far more than the per-minute price does.

GPT-Live-1 hands work across an asynchronous boundary to a backend named in the session configuration. O​penAI's published WebRTC example wires that backend to GPT-5.6 Terra through the Responses API. The voice layer keeps the conversation going while the backend thinks, and folds the answer back in when it arrives. The consequence is that the reasoning half is an ordinary token-metered text call that you choose, price and swap independently. The voice half is not portable at all: there is one model ID, gpt-live-1, one endpoint at v1/live/sessions, WebRTC rather than the WebSocket transport most Realtime integrations were built on, and no path through Chat Completions, Responses, Batch or the fine-tuning endpoints. Rate limits are denominated in simultaneous sessions — 25 on Tier 1, rising through 50, 200, 300 and 500 — which tells you O​penAI thinks the unit of scale is a phone line rather than a request.

Grok Voice 3.0 is narrower and cheaper to adopt. It runs over the same realtime WebSocket endpoint as Grok Voice 2.0, with the same session parameters, the same event names and the same bearer-token handshake, and tool use goes through a tools array in the session update the way it did before. There is no transport migration. There is also no documented backend-delegation pattern of the sort GPT-Live-1 publishes, which means the reasoning you need has to be reachable as a tool the voice model calls synchronously rather than as a background job it supervises.

That difference decides more deployments than the index does. A team already on xAI's realtime API moves to Grok Voice 3.0 by changing a model string — or by doing nothing, because grok-voice-latest now resolves to it and existing alias deployments were moved automatically. A team on O​penAI's older Realtime models moving to GPT-Live-1 is looking at a transport rewrite plus a delegation rewrite.

Line by line

Every figure below is Artificial Analysis's own measurement dated 13 October 2026 unless the line says otherwise.

• Speech to Speech Index — Grok Voice 3.0 86.9 vs GPT-Live-1 (Astra, medium) 83.2. There is also a lower GPT-Live-1 configuration measured at 80.1 in its Sol low-effort variant, and the index gap depends entirely on which one you mean.

• Agentic performance, τ-Voice — GPT-Live-1 Astra medium 74.5% vs Grok Voice 3.0 71.6%. O​penAI leads this one, and leads it on the component that predicts whether a caller's problem is resolved.

• Speech reasoning, Big Bench Audio — Grok Voice 3.0 98.5% vs GPT-Live-1 Astra medium, which sits in the same band. Both are close enough to the ceiling that this axis has stopped separating anyone.

• Arena preference — Grok Voice 3.0 holds the highest Elo on the board at 1070. GPT-Live-1 does not hold the arena in any configuration.

• Time to first audio — Grok Voice 3.0 0.81 seconds vs GPT-Live-1 Astra medium at 1.34 seconds. Roughly 40% quicker to start speaking, and the most concrete performance claim either vendor makes.

• Cost per completed task, Big Bench Audio — Grok Voice 3.0 $0.025 vs GPT-Live-1 Astra medium $0.030. Cheap and accurate per task is the combination that ends procurement debates.

• Price per minute of audio — GPT-Live-1 $0.05, billed by the second, with no rounding up of partial minutes and a 15-second session-initialisation charge that is credited against later duration rather than stacked on top; Grok Voice 3.0 $0.08 per minute plus $0.004 per text input.

• Multimodal input — GPT-Live-1 takes audio and text and its documentation lists image and video as explicitly unsupported; Grok Voice 3.0 has no visual grounding either. Neither of these two can see, and neither vendor has published a date for changing that.

• Voices — GPT-Live-1 ships 12; Grok Voice 3.0 keeps xAI's built-in set and the custom-voice endpoint that trains a voice from a reference clip of up to 120 seconds.

The bill, for a workload you can picture

Ten thousand minutes of conversation a day across thirty days is 300,000 minutes a month.

• GPT-Live-1 — $0.05 per minute of voice, so $15,000 a month on the voice layer alone.

• Grok Voice 3.0 — $0.08 per minute, so $24,000 a month, plus $0.004 per text input.

Both numbers are only the voice half. GPT-Live-1's backend reasoning, its tool calls and any web search bill separately at the backend model's own rates, and a session pointed at a frontier reasoner that searches on every turn can cost more in delegation than in audio. Grok Voice 3.0's synchronous tool calls run on whatever backend answers them, under the same rule. In both cases the "voice model price" is a fraction of the call, and the fraction depends on how hard your callers think.

That is the layer where a router is actually useful, and it is worth being exact about the boundary. OrcaRouter serves neither realtime endpoint — gpt-live-1 lives behind O​penAI's Live Sessions API and Grok Voice 3.0 behind xAI's own, and nothing can interpose on a duplex audio session that belongs to one vendor. What we do run is the delegation half. GPT-5.6 Terra, the exact backend in OpenAI's own published example, is served through OrcaRouter at O​penAI's list price of $2.00 per million input tokens and $12.00 per million output with both chat and Responses endpoints exposed, alongside 200-plus other models on one key, at provider list price with zero markup and automatic failover when a provider degrades. Because nothing is marked up, a vendor price cut on the reasoning tier is live on our side the same day — which matters in a category where one flagship raised its per-minute rate 60% inside three months while another held it flat.

What each vendor says that no board covers

Both vendors publish numbers about their own production traffic. Neither is independently verifiable, and the honest way to read them is as evidence of confidence rather than of capability.

O​penAI's figures, published with the API release and reproduced by nobody: full-duplex interactivity at 80.1% against 45.4% for GPT-Realtime-2.1, turn-taking latency cut from 1.4 seconds to 0.8, tool-calling accuracy from 60% to 87%. Yelp is cited on improved phone-based reservation handling, and language-learning platform Speak reported close to 80% fewer interruptions than its previous turn-based system — self-reported by a company with an interest in the result, and still the most concrete deployment figure in the release.

SpaceXAI's figures, published with its launch and equally unreproduced: Grok Voice 3.0 powers millions of daily conversations across Tesla vehicles and SpaceX, Starlink call centres run 20,000 customer calls a day in multiple languages with roughly 70% resolved without a human and about one in four sales calls ending in a purchase, and a rounded 74% on ServiceNow's EVA-Bench. Treat the containment rate as a system-level number wearing a model-level costume — it reflects prompt design, retrieval and escalation policy as much as the model, and no external party can separate them.

Which one to actually pick

Take Grok Voice 3.0 when latency is the constraint and the stack is already on xAI. Time to first audio of 0.81 seconds is the fastest published figure among frontier voice models, the WebSocket path means no transport rewrite, and the migration cost is a model string. Take it also if your callers judge the experience rather than the outcome — it holds the arena on preference and the composite on quality, and a caller who enjoys the conversation hangs up happier even when the task escalates.

Take GPT-Live-1 when the task is the product. It leads τ-Voice at 74.5% against 71.6%, its delegation pattern is documented and its reasoning backend is a portable token-metered call rather than an opaque tool, and per completed task it is competitive on cost despite the index gap. Take it also if you are building from scratch and have no WebSocket integration to preserve, because the transport story only cuts against you when you already have one.

The one answer to reject is picking on the composite. A 3.7-point index gap and a 2.9-point τ-Voice gap point in opposite directions, and the cost difference points in a third — GPT-Live-1 is 37.5% cheaper per minute of audio before you count text input. Anyone choosing between these two on 86.9 versus 83.2 is choosing on the number least connected to whether their phone line works.

What would settle it: a general-availability commitment and a published latency figure from O​penAI for the Astra configuration that holds the top of the board, since the 0.81-versus-1.34 comparison currently rests on xAI's number against a configuration O​penAI's own docs treat as an example; or a price move from either side. The rate card is the one part of this matchup that has moved before, and it moved upward.

Comparisons like this one are also where the failure modes show up: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.