GPT-Live-1 vs Eleven v4: The Billing Unit Decides Before the Voice Does

Put GPT-Live-1 and Eleven v4 side by side and the first difference you meet is not a benchmark. It is a unit. GPT-Live-1 charges $0.05 for every minute the session is open. Eleven v4 charges $0.08 for every thousand characters the agent actually says. One meter runs on the clock, the other on the script, and which one you should buy follows from that before either model produces a syllable.

That distinction gets lost in most comparisons, which tend to park a speech-to-speech index next to a text-to-speech Elo and call it a head-to-head. It isn't. Here is what the two products actually are, what the numbers that exist can and cannot tell you, and the arithmetic that resolves it for a support line, an audiobook, and the awkward cases in between.

Two products, two jobs

GPT-Live-1 is OpenAI's full-duplex speech model, on the API since 10 September 2026. It listens while it speaks, decides many times a second whether to yield the floor, and delegates reasoning to a separate text model when a question needs it. It has exactly one endpoint — v1/live/sessions — and no Chat Completions or Realtime surface, which means it is not a model id you swap into an existing Realtime integration; it is a session you open.

Eleven v4 is a text-to-speech model, released 28 September 2026 alongside a low-latency v4 Turbo. You hand it text, it returns audio in one of 90-plus languages, with 10-second instant voice cloning, inline audio tags for emotion and pacing, and a 10,000-character request ceiling. It has no ears. It cannot hear the caller, notice that they interrupted, or decide to stop because the sentence landed badly.

That asymmetry is the whole story. A voice agent built on Eleven v4 is a pipeline: transcribe, think, speak. A voice agent built on GPT-Live-1 is one model with a text model attached as a consultant. The pipeline is cheaper and more controllable. The end-to-end model is more natural and much harder to interrupt. Those are different products for different rooms, not competing implementations of the same one.

What the leaderboards can honestly say

There is no single board where these two compete, and anyone showing you one is comparing unlike scores. On Artificial Analysis' Provider Voice Arena — native voices, blind preference, captured 29 October 2026 — Eleven v4 sits at 1334 Elo, second of the field, with v4 Turbo at 1336 taking first. Nobody disputes that the voice quality is excellent: 91.7% pronunciation robustness in that lab's measurement, the strongest of the major commercial models at launch.

GPT-Live-1 does not appear on that board at all, because it does not generate text-to-speech in the sense the board measures. It appears on the Speech Agent Arena, which scores blind preference over complete booking-style conversations. There, GPT-Live-1 (Astra, medium) sits at 1037 Elo with an 87.4% task success rate, and the Sol-low configuration at 1041 with 90.9%. On the Speech-to-Speech Index — a weighted blend of reasoning, agentic performance, arena preference and task success — it scores 83.2, second of 21 scored models.

The comparison that actually holds is GPT-Live-1 against Eleven's conversational products, not its TTS model. On the same arena, Eleven Realtime v1 Max scores 1003 with 87.1% task success, and the ElevenLabs Agents default cascade scores 989 with 90.5%. So the end-to-end model leads the realtime alternative by roughly 35–50 Elo, and the cascaded agent stack sits about 50 points lower with a slightly better success rate. Those numbers are close enough that a single deployment's telephony, prompt design and background noise will decide the outcome, not the board.

One more honest caveat: the GPT-Live-1 figures carry a configuration label. Its reasoning is done by a named backend, and the Astra-medium and Sol-low configurations score differently because they delegate differently. Eleven's numbers carry a product label instead. Neither is a controlled experiment against the other.

The arithmetic, on a three-minute call

Take an inbound support call where the agent speaks 900 characters a minute — a brisk but ordinary pace — and the caller speaks for about 40% of the three minutes.

• GPT-Live-1 — three minutes at $0.05 comes to $0.15, plus the backend text model for whatever reasoning the call required. Per minute: $0.050.

• Eleven v4 — the agent speaks roughly 1,620 characters, so 1.62 × $0.08 = $0.13, plus transcription and your own language model. Per minute: about $0.043 before STT and LLM.

• Eleven v4 Turbo — the same 1,620 characters at $0.04 per thousand: $0.065, or roughly $0.022 per minute before STT and LLM.

Turbo is cheaper than GPT-Live-1 even after you add a speech-to-text leg and a modest language model, which is why the cascade keeps winning cost reviews. But the comparison changes shape the moment the caller talks more than the agent. Every second the human monologues is a second Eleven v4 charges nothing for and GPT-Live-1 charges full rate — including the silence while someone reads a policy number off a card, and including the hold while your backend thinks.

That is not a rounding difference. On a line where the agent speaks 30% of the time, the TTS bill drops by two thirds while the per-minute bill does not move at all. The flat minute is a bet that your callers will talk, and it pays off exactly when they do.

The arithmetic, on a forty-minute narration

Now the same two models on an audiobook chapter — 36,000 characters, one speaker, no interruptions, no reasoning, deterministic output required.

• Eleven v4 — 36 × $0.08 = $2.88 for the chapter, and you can regenerate a single paragraph for nine cents when the pronunciation is wrong.

• GPT-Live-1 — 40 minutes at $0.05 = $2.00, and you cannot regenerate a paragraph, because the unit of work is a live session rather than a piece of text.

GPT-Live-1 looks cheaper and is the wrong tool. It has no documented way to take a 36,000-character script and return a deterministic file; it is a conversation surface, and a conversation with a book is not a narration. The honest verdict on this row is not "Live costs less", it is "this model does not do this job".

Where the per-minute model actually earns its premium

Three situations justify it, and they share one property: the human holds the floor unpredictable amounts of the time.

• Interruption-heavy intake — insurance claims, clinical triage, anything where the caller corrects themselves mid-sentence. A cascade has to detect the interruption, stop playback, and re-enter the loop; GPT-Live-1 is already listening.

• Booking and dispatch — short turns, rapid back-and-forth, heavy tool calling. OpenAI's own figures claim 87% tool-calling accuracy on this model against 60% for the Realtime generation it replaces, and those are vendor-reported, not independently reproduced.

• Anywhere latency is the product — a 0.8-second turn against a 1.4-second turn is the difference between a conversation and a walkie-talkie. Also vendor-reported.

And three situations argue against it:

• Narration, read-aloud, audiobooks, video voiceover — text in, file out, no ear required.

• Brand voice — Eleven v4's cloning is a product feature with a retraining caveat; GPT-Live-1 uses fixed model voices, so "our founder's voice" is not on the menu.

• Long silences — hold music, muted lines, agents waiting on a slow backend. Every one of those seconds bills.

Migration is not symmetric either

Moving to Eleven v4 from v3 is a model-id change plus a caveat: existing cloned voices need retraining, and SSML is replaced by inline audio tags, so anything that generated SSML has to be rewritten. Annoying, but contained inside your TTS layer.

Moving to GPT-Live-1 from a Realtime integration is a rebuild. Different endpoint, different session semantics, different billing unit, and no compatibility shim. Budget for the orchestration layer, not just the swap — and note that most third-party voice-agent platforms still do not list GPT-Live-1 as a selectable provider, so you are likely writing the telephony glue yourself.

What routing changes about this decision

Whichever leg you pick, the decision is not all-or-nothing across your stack. The voice layer is the one piece that stays vendor-specific: OrcaRouter does not serve GPT-Live-1, and a speech-synthesis model like Eleven v4 is a separate contract wherever you buy it. What sits behind the voice layer is a different matter — the reasoning calls, the tool-calling model, the fallback path when a provider degrades all run through one key on OrcaRouter, across 200-plus models including GPT-6 Astra, GPT-5.6 Terra and GPT-5.6 Sol, the tiers GPT-Live-1 delegates to.

That split is the practical version of this comparison. Rent the ear from whoever has the best one this quarter, and keep the brain routable so that a price cut, an outage, or a better model does not require touching the part your users can hear. Provider list prices pass through at 0% markup, so a vendor cut on the reasoning tier shows up on your bill the same day it is announced.

Pick by the shape of the conversation

If your users interrupt, correct themselves, and expect the agent to know when to stop talking, GPT-Live-1 at $0.05 a minute is the better product and probably the better buy — the premium buys behaviour you cannot get from a pipeline at any price. If your users are listening rather than talking, Eleven v4 or v4 Turbo is cheaper per delivered word, more controllable, clonable to your brand, and the correct answer; the 1334 Elo is a bonus on top of that.

The failure mode to avoid is picking by leaderboard position. v4's 1334 and Live's 83.2 measure different things on different scales, and a team that chooses between them on those two numbers is choosing between a ruler and a thermometer.