● Cheapest per minute Gemini Live at about $0.023 combined - and free of charge on the free tier.
● Simplest to forecast GPT-Live-1 at a flat $0.05 per minute, billed per second with no rounding up.
● Lowest cost per task, per xAI Grok Voice 3.0 at $0.08 per minute - the most expensive by the clock.
A three-fold price gap that may not survive contact with your actual call length.
Voice APIs are priced in two incompatible ways. Two of these three bill audio by the minute, one bills it by the token and converts for you, and the cheapest by the clock is not necessarily the cheapest by the job. Here is what each actually charges.
The rates, as published
| Grok Voice 3.0 | GPT-Live-1 | Gemini Live |
|---|
| Headline rate | $0.08/min ($4.80/hr) | $0.05/min | $3.00/1M audio in, $12.00/1M audio out |
| Converted per minute | $0.08 | $0.05 | ~$0.023 ($0.005 in + $0.018 out) |
| Billing granularity | Per minute | Per second, not rounded up | Per token |
| Text input | $0.004 per text input | Backend model billed separately | Separate text rates apply |
| Free tier | Not stated | Not supported | Free of charge |
| Rate limiting | Not stated on the pricing page | Concurrent sessions: Build 50, Launch 300, Grow 500 | Standard API quotas |
| Models | grok-voice-3.0, grok-voice-latest | gpt-live-1 | gemini-3.8-live, gemini-3.1-flash-live-preview |
The headline rate is not the bill for any of them. GPT-Live-1 bills backend model and tool usage separately at the configured model’s normal rates, so a voice agent that calls functions costs the $0.05 plus whatever the reasoning model charges. Grok adds $0.004 per text input. Gemini charges audio in and out at different rates, so a chatty assistant costs more per minute than a listening one.
Why the ranking inverts on cost per task
xAI publishes a comparison that cuts against its own per-minute price. On Big Bench Audio, it reports average cost per task of $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).
If that holds on your workload, the most expensive model per minute is the cheapest per job, because it finishes sooner. That is a coherent claim - a voice agent that resolves a call in 90 seconds at $0.08 costs less than one that takes four minutes at $0.05.
Treat it as a vendor-published comparison, because it is one. xAI chose the benchmark, the competing configurations and the effort settings. The direction of the argument is sound and the specific numbers are not independently verified. The honest version: per-minute price tells you almost nothing until you know your average call length.
Grok Voice 3.0
Verdict: the strongest benchmark sheet of the three, at the highest clock rate.
Released 14 October 2026 at $0.08 per minute, unchanged from Voice 2.0. xAI reports it first on the Artificial Analysis Speech-to-Speech Index at 86.9, up from 81.3 for the previous version.
- Full Duplex Bench, covering pauses, turn-taking and interruptions: 97%
- Big Bench Audio, speech reasoning: 98.5%
- Task success rate: 94.8%
- EVA-Bench: 74%, which xAI reports as the highest recorded
- Latency under one second, with τ-Voice latency 40% below GPT-Live-1
It powers voice mode in Grok Bot and handles conversations in Tesla vehicles. xAI also says it runs Starlink call centres taking 20,000 calls a day across multiple languages, with about 70% resolved without a human - which is the most useful data point on the page, because it is a production deployment rather than a benchmark.
Honest limitation: every figure above is xAI’s own, and the pricing page states no free tier and no concurrency limits, so you cannot plan capacity from the public documentation.
Best for: support and call deflection, where resolving the call matters more than the per-minute rate.
GPT-Live-1
Verdict: the easiest of the three to forecast, and the only one that publishes concurrency limits.
A flat $0.05 per minute, billed per second, and OpenAI states session time is not rounded up to the next whole minute. For short interactions that detail is worth more than it sounds: at per-minute rounding, a 20-second call costs a full minute.
Rate limits are expressed as concurrent sessions rather than requests - 50 on Build, 300 on Launch, 500 on Grow. That is the number you size a call centre against, and the other two providers do not publish an equivalent. There is no free tier.
Streaming and function calling are supported. Structured outputs, fine-tuning, predicted outputs, image and video are not. Knowledge cutoff is 31 July 2025.
Honest limitation: the $0.05 covers the voice session only. Backend model and tool usage bills separately at the configured model’s normal rates, so an agent doing real work costs meaningfully more than the headline.
Best for: teams that need to size concurrency and commit to a number before launch.
Gemini Live
Verdict: less than half the price of the others, and free while you build.
Google prices Live by token rather than by minute: $3.00 per million audio input tokens and $12.00 per million audio output, which Google converts to roughly $0.005 and $0.018 per minute. Combined, that is about $0.023 a minute for a balanced conversation - under half GPT-Live-1 and under a third of Grok.
The asymmetry matters: output costs four times input. An assistant that talks a lot costs far more than one that mostly listens, which is the opposite of how a flat per-minute rate behaves.
The free tier is the real story. The Live models - 3.8 Live, 3.8 Live Extended Thinking and 3.1 Flash Live Preview - are free of charge for input and output on the free tier. Google notes free-tier usage may be used to improve its products, which rules it out for anything confidential but makes it the obvious place to prototype.
Speech-to-speech translation is priced separately at $3.50 in and $21.00 out per million.
Honest limitation: token-denominated audio pricing is harder to forecast than a per-minute rate, because token count varies with speech density and language.
Best for: prototyping at zero cost, and for production workloads where the assistant listens more than it speaks.
Where ElevenLabs fits, and where it does not
ElevenLabs is often named alongside these three, but it is solving a different problem. It is text to speech, not speech to speech - you send text and get audio. There is no listening half, so it is not a drop-in alternative for a voice agent that holds a conversation.
Its API prices per thousand characters rather than per minute: $0.08 per 1K characters for v4, v3 and v2 Multilingual, and $0.04 for the Turbo-class models - v4 Turbo, v3 Conversational and Flash/Turbo. Plans run from Starter at $6 a month for 75,000 v4 characters up to Business at $990 for 12,375,000, with roughly double those allowances on the Turbo-tier models.
Use it for narration, audiobooks, IVR prompts and anything where you control the script. Do not use it for the interactive half of a voice agent - pair it with a separate speech-to-text and reasoning stack, or use one of the three speech-to-speech models above.
What a real month costs
Take a support line handling 1,000 calls a month. The answer changes entirely with average call length:
| Average call | Grok Voice 3.0 | GPT-Live-1 | Gemini Live |
|---|
| 2 minutes | $160 | $100 | $46 |
| 4 minutes | $320 | $200 | $92 |
| 6 minutes | $480 | $300 | $184 |
Those are clock-rate figures only, and they exclude backend model costs on GPT-Live-1. If xAI’s per-task claim holds and Grok resolves the same calls in half the time, the two-minute Grok column competes with the four-minute GPT-Live column. That is the entire argument, and it is testable on your own call recordings in an afternoon.
Decision framework
- You are prototyping → Gemini Live, free tier, as long as the audio is not confidential.
- You need to size concurrency before launch → GPT-Live-1, the only one publishing session limits.
- Your calls are short → GPT-Live-1, because per-second billing with no rounding stops you paying for a minute you did not use.
- Your assistant listens more than it talks → Gemini Live, where input is a quarter the price of output.
- You are deflecting support calls → test Grok Voice 3.0 against your own recordings. If it resolves faster, the higher rate pays for itself.
- You only need narration from a script → none of these. ElevenLabs Turbo at $0.04 per 1K characters.
FAQ
Which voice API is cheapest?
Gemini Live, at roughly $0.023 per minute combined, and free on the free tier. GPT-Live-1 is $0.05 and Grok Voice 3.0 is $0.08.
Does Gemini Live really have a free tier?
Yes. Google lists the Live models as free of charge for input and output on the free tier, and notes that free-tier usage may be used to improve Google products. GPT-Live-1 explicitly does not support a free tier.
Is Grok Voice 3.0 worth the higher price?
Only if it finishes calls faster on your workload. xAI reports $0.025 average cost per task on Big Bench Audio against $0.030 for GPT-Live-1, which would make it cheaper per job despite costing more per minute. That is a vendor-published figure and worth testing yourself.
Why is the Gemini price given per million tokens?
Google bills audio by token rather than by minute, then publishes an approximate per-minute conversion. Token count varies with speech density, so the converted rate is an estimate rather than a fixed price.
Can I use ElevenLabs for a voice agent?
Not on its own. ElevenLabs is text to speech - it has no listening half. You would need separate speech-to-text and reasoning layers, which usually costs more than using one speech-to-speech model.
Does the headline rate cover everything?
No. GPT-Live-1 bills backend model and tool usage separately at normal rates. Grok adds $0.004 per text input. Gemini charges output at four times input.
Related
Sources
Every rate was read from the vendor’s own documentation on 2 November 2026. Benchmark scores attributed to xAI are xAI figures and have not been independently validated. Voice pricing changes often - check the linked page before committing.