GPT-Live-1 vs Voxtral 3: The Price That Isn't the Price

Voxtral 3 advertises a complete voice agent at $0.04 a minute. GPT-Live-1 advertises $0.05 a minute and then bills you again for the thinking. On the headline numbers Mistral is 20% cheaper and includes more — speech-to-text, text-to-speech and a reasoning model inside one rate — while OpenAI's number covers only the voice layer. That reading is technically correct and practically misleading, because the two rates are not buying the same thing.

Both shipped within six weeks of each other: GPT-Live-1 to the API on 10 September 2026, Voxtral 3 in public preview on 20 October. Both target the same customer — someone building a phone agent that has to hold a conversation. They get there by opposite routes, and the cost difference only survives if you never interrupt them.

What is actually inside each minute

Start with what the meters include, because this is where the comparison turns.

• Voxtral 3 full agent — $0.04 per minute with Mistral Small 4 as the reasoner, $0.08 with a stronger one. That covers transcription, synthesis and the language model.

• GPT-Live-1 — $0.05 per minute for voice only, billed by the second. Backend reasoning, tool calls and web search bill separately at normal model rates.

• Voxtral 3 TTS alone — $13 per million characters, roughly $0.70 per hour of speech at a normal pace.

• GPT-Live-1 input audio — measured at $5.83 per hour of input audio on its high-reasoning configuration, $4.47 on its low one, per Artificial Analysis' fixed 40-question set.

So the honest per-minute comparison on a ten-minute call is $0.40 for Voxtral 3 all-in against $0.50 plus backend for GPT-Live-1 — call it $0.50 to $0.55 depending on how much the agent has to think. Mistral wins that arithmetic, and it wins it by including a component OpenAI leaves out.

Unless the call is mostly silence. Voxtral's voice-agent rate is metered like a session; GPT-Live-1's certainly is. Neither vendor publishes a silence exemption, and any deployment that parks callers on hold should assume both meters keep running. That difference does not favour either side, but it does mean the headline rates are the wrong place to optimise.

Pipeline against duplex, again

Voxtral 3 is a stack, not a model: transcription models (Mini Transcribe 3 at 4B parameters, Small, and a realtime variant), a synthesis model, and a LiveKit-based Voice Agent SDK that wires them together with a language model in the middle. Mistral's own latency claim is about 90 milliseconds to first audio, which is genuinely fast for the synthesis leg.

Fast synthesis is not the same as good turn-taking. A cascade still has to decide when the human finished speaking, and that decision is made by a voice-activity detector or a turn-detection model sitting in front of the pipeline. When it guesses wrong, the agent either talks over the caller or waits a beat too long, and no amount of 90ms synthesis fixes that.

GPT-Live-1 is one network that hears and speaks at once, deciding many times per second whether to yield. That is the capability you are paying the extra cent for, and it is not a feature you can assemble from parts. It is also why OpenAI's model has no transcription or synthesis line items: there is nothing to itemise.

The practical test is not latency but interruption rate. If your callers talk in complete sentences and wait their turn, a pipeline is fine and cheaper. If they say "no wait, I meant the other one" halfway through, you are buying duplex behaviour whether or not you have admitted it.

Languages, dialects and where the data sits

Mistral ships Voxtral 3 with 35-plus languages and 50-plus dialects, emotion and style steering, zero-shot cloning plus a "Pro Voice" adaptation path that takes about 30 minutes of reference audio. GPT-Live-1's documentation is silent on dialect coverage and offers fixed model voices with no cloning.

If your product is European — French, German, Italian, Spanish, Dutch — Voxtral 3 is the more complete offer, and Mistral's EU positioning is a real argument in procurement conversations that OpenAI's does not answer as directly. That is a business fact rather than a benchmark, and for a lot of buyers it decides the deal before latency is discussed.

If your product is English-language support or booking in a single market, the language breadth is inventory you will not use, and the duplex behaviour is the thing you will notice on every call.

The open-weights trap

Worth flagging because it catches people: Voxtral 3 itself is a closed API. The downloadable model in this family is an older 4B TTS checkpoint from March 2026, released under CC BY-NC 4.0 — non-commercial — a different, older model with a lower arena score. Enterprise self-deployment of Voxtral 3 is a negotiated arrangement, not a download link.

So neither model in this comparison gives you weights. GPT-Live-1 has none by design. Voxtral 3 has none by policy. If owning the model is a requirement, both are out and the conversation moves to Fish Audio Drama 3 or Ming-3.0-TTS instead.

What the boards measure, and what they don't

Voxtral 3 TTS sits at 1223 Elo on Artificial Analysis' Provider Voice Arena, captured 29 October 2026 — fourteenth of the field, at $13 per million characters. That is a solid synthesis result: behind Eleven v4 at 1334, ahead of several models costing more.

GPT-Live-1 does not appear on that board, and anyone who places its Speech-to-Speech Index score of 83.2 next to Voxtral's 1223 is comparing a weighted conversation index with a blind-preference Elo. On the boards where conversational systems do appear, GPT-Live-1 scores 1037 to 1041 Elo on the Speech Agent Arena with 87.4% to 90.9% task success. Voxtral 3's agent stack has not yet accumulated enough arena conversations to be listed.

That absence is the most useful fact in this section. There is no independent head-to-head between these two, which means every comparison you read — including this one — is reasoning from published rates and architecture, not from a controlled experiment. Anyone who tells you one is measurably more natural on the phone is quoting a vendor.

Where the routing argument lands

Neither model is served by OrcaRouter. Voxtral 3 is Mistral's API; GPT-Live-1 is OpenAI's. What is routable is the reasoner each one delegates to — and here the two architectures diverge in a way that matters for how you architect the stack.

Voxtral 3 lets you choose the reasoning model inside its own agent rate: Small 4 at $0.04, something stronger at $0.08. GPT-Live-1 leaves the backend open entirely, which means you can point it at whatever text model you like — including GPT-6 Astra, GPT-5.6 Terra or a cheaper tier, behind one key, with the routing DSL letting you pin different models for different task types inside the same deployment.

That flexibility is worth real money on a mixed workload: a cheap model handles address confirmation and a frontier model handles the disputed charge, without a second contract or a second integration. Provider list prices pass through at 0% markup, so a cut on the reasoning tier reaches your bill the day it lands rather than at the next contract renewal.

The decision

Pick Voxtral 3 if you are building in Europe, need many languages and dialects, want one invoice that covers transcription through synthesis, and your callers are reasonably polite about taking turns. It is cheaper per minute all-in, it is the more complete stack, and the 1223 synthesis score is respectable.

Pick GPT-Live-1 if interruptions are the product — intake, triage, dispatch, anything where the caller corrects themselves and the agent has to hear it happening. You will pay more per minute, you will own less, and you will get the one behaviour a pipeline cannot buy at any price.

The trap to avoid is comparing the two headline rates and concluding that $0.04 beats $0.05. Those numbers are not measuring the same minute: one includes a transcriber, a synthesiser and a reasoner; the other includes an ear. Buy the one whose contents match the job, then argue about the price.